Method and apparatus for channel estimation using auto-encoder in communication system

Abstract

A method of a UE according to the present invention may comprise the steps of: receiving, from a base station, a CFI transmission period and an LVD of an encoder that performs online training; on the basis of the CFI transmission period and the LVD, determining an encoder for executing online training; receiving the first RS from the base station; generating a CFI by compressing the received first RS via the determined encoder; and transmitting the first CFI to the base station on the basis of the CFI period.

Description

Technical Field

[0001] The present invention relates to enhanced communication technology, and more particularly, to technology for channel estimation. Background Art

[0002] Communication networks (e.g., 5G communication networks or 6G communication networks) are being developed to provide enhanced communication services compared to existing communication networks (e.g., long term evolution (LTE), LTE-Advanced (LTE-A), etc.). 5G communication networks (e.g., New Radio (NR) communication networks) can support frequency bands below 6 GHz and frequency bands above 6 GHz. In other words, 5G communication networks can support frequency region 1 (FR1) bands and / or FR2 bands. Compared with LTE communication networks, 5G communication networks can support various communication services and scenarios. For example, usage scenarios of 5G communication networks may include enhanced Mobile BroadBand (eMBB), ultra-reliable low-latency communication (URLLC), massive machine type communication (mMTC), etc.

[0003] Compared to 5G communication networks, 6G communication networks can support a wide variety of communication services and scenarios. They can meet requirements for ultra-performance, ultra-bandwidth, ultra-space, ultra-precision, ultra-intelligence, and / or ultra-reliability. They can support multiple broadband frequencies and be applied to a variety of use cases, such as terrestrial communications, non-terrestrial communications, and sidelink communications.

[0004] On the other hand, the current 5G NR adopts a codebook-based channel state information (CSI) feedback scheme for downlink CSI estimation. The codebook-based CSI feedback scheme refers to a scheme in which the base station periodically or aperiodically sends a CSI reference signal (CSI-Reference Signal, CSI-RS) to the UE for downlink channel state estimation, and the UE reports the CSI to the base station. In this scheme, the CSI may include a channel quality indicator (CQI), a precoding matrix indicator (PMI) and a rank indication (RI). The base station can estimate the CSI using a predefined codebook based on the CQI, PMI and RI reported by the UE.

[0005] 3GPP is currently discussing artificial intelligence / machine learning (AI / ML) solutions for the physical layer as part of a study item (SI) in Release 18. AI / ML solutions can involve training a model using output data corresponding to input data to extract specific patterns or parameters, and predicting output data for arbitrary input data based on the trained model. Within 3GPP, discussions are ongoing regarding the application of AI / ML solutions in technical areas such as CSI feedback, beam management, and positioning accuracy enhancement.

[0006] However, the solutions discussed or agreed upon in 3GPP to date are only high-level suggestions for improving AI / ML performance, and have not yet addressed specific procedures for utilizing these technologies. Therefore, in order to support AI / ML-based CSI feedback schemes as an alternative to existing codebook-based CSI feedback schemes that utilize CQI, PMI, and RI, it is necessary to define new CSI feedback procedures and corresponding parameters that more accurately consider the characteristics of AI / ML, beyond what has been proposed or is currently under discussion. Summary of the Invention

[0007] Technical issues

[0008] The present invention is directed to providing a method and apparatus for CSI feedback for applying AI / ML in a communication system.

[0009] Technical Solution

[0010] A method of a user equipment (UE) for achieving the above-mentioned purpose according to an exemplary embodiment of the present invention may include: receiving a channel characteristic indicator (CFI) transmission period and a latent variable dimension (LVD) of an encoder that performs online learning from a base station; determining a first encoder that performs online learning based on the CFI transmission period and the LVD; receiving a first reference signal (RS) from the base station; generating a first CFI by compressing the first RS received by the first encoder; and sending the first CFI to the base station based on the CFI transmission period, wherein the first CFI is determined based on the product of the number of nodes in the latent space of the first encoder and the LVD.

[0011] The LVD may correspond to the number of bits required to represent each latent variable value included in the latent space of the first encoder.

[0012] The CFI transmission period may be determined based on at least one of a network size or a channel coherence time of the first encoder.

[0013] The LVD may be determined based on at least one of a channel resolution or a latency requirement of the base station.

[0014] The method may further include: receiving retraining-related information of the first encoder from the base station; reconfiguring the first encoder based on the retraining-related information; receiving a second RS from the base station; generating a second CFI by compressing the received second RS by the reconfigured first encoder; and sending the second CFI to the base station.

[0015] The retraining-related information may include at least one of information indicating retraining of the encoder, information on a dropped node of the encoder, or identifier information of the first encoder, and the identifier information of the first encoder may further include a drop count.

[0016] The method may further include: receiving training termination information indicating the termination of training of the first encoder from the base station; updating the first encoder based on information obtained by training the first encoder before receiving the training termination information; receiving a third RS from the base station; generating a third CFI by compressing the third RS by the updated first encoder; and sending the third CFI to the base station.

[0017] The method may further include: receiving training expiration information including optimal encoder information from a base station; updating a first encoder based on the optimal encoder information from the base station; configuring the updated first encoder as a channel estimation encoder; receiving a fourth RS from the base station; generating a fourth CFI by compressing the fourth RS by the updated first encoder; and sending the fourth CFI to the base station, wherein the optimal encoder information includes an encoder identifier and information about the number of discards of the encoder.

[0018] According to an exemplary embodiment of the present invention, a method of a base station for achieving the above-mentioned purpose may include: determining a channel characteristic indicator (CFI) transmission period and a latent variable dimension (LVD) of an encoder that performs online learning with a user equipment (UE); sending the CFI transmission period and LVD to the UE; receiving a first CFI from the UE based on the CFI transmission period; and recovering a received reference signal (RS) from the received first CFI through a decoder.

[0019] The method may further include: obtaining a first reference signal received power (RSRP) value of the recovered received RS; checking whether the first RSRP value is saturated; based on the saturation of the first RSRP value, checking whether the first RSRP value is equal to or greater than a preset threshold; and based on the first RSRP value being equal to or greater than the preset threshold, sending training termination information of the encoder to the UE, wherein the first CFI is determined based on the product of the number of nodes in the latent space of the encoder performing online learning and the LVD.

[0020] The first RSRP value may be determined to be saturated based on the first RSRP value being equal to an RSRP value obtained from a CFI previously received by the UE or being within a predetermined range of an RSRP value obtained from a CFI previously received by the UE.

[0021] The method may further include: generating encoder retraining related information based on the first RSRP value being unsaturated; sending the encoder retraining related information to the UE; and performing the encoder retraining process using the UE.

[0022] The retraining-related information may include at least one of information indicating retraining of the encoder, information on a dropped node of the encoder, or identifier information of the encoder, and the identifier information of the encoder may further include a drop count.

[0023] The method may further include: checking the number of discards corresponding to the RSRP having the highest RSRP value among the trained RSRP values ​​based on the retraining not being completed within a predetermined time; and sending identifier information of the encoder including the number of discards and training expiration information to the UE.

[0024] The CFI transmission period may be determined based on at least one of a network size of an encoder performing online training or a channel coherence time.

[0025] The LVD may be determined based on at least one of a channel resolution or a latency requirement of the base station.

[0026] The method may further include: transmitting a second RS to the UE after training is completed; receiving a second CFI corresponding to the second RS from the UE; obtaining a second RSRP value of the received RS from the second CFI; and estimating a downlink channel toward the UE using the second RSRP value.

[0027] According to an exemplary embodiment of the present invention, a user equipment (UE) may include: a transceiver configured to send signals to and receive signals from a base station; and at least one processor, wherein the at least one processor may enable the UE to perform: receiving a channel characteristic indicator (CFI) transmission period and a latent variable dimension (LVD) of an encoder that performs online learning from the base station; determining a first encoder that performs online learning based on the CFI transmission period and the LVD; receiving a first reference signal (RS) from the base station; generating a first CFI by compressing the first RS received by the first encoder; and sending the first CFI to the base station based on the CFI transmission period, wherein the first CFI may be determined based on the product of the number of nodes in the latent space of the first encoder and the LVD.

[0028] The CFI transmission period may be determined based on at least one of a network size or a channel coherence time of the first encoder, and the LVD may be determined based on at least one of a channel resolution or a latency requirement of the base station.

[0029] At least one processor may further cause the UE to perform: receiving retraining-related information of the first encoder from the base station; reconfiguring the first encoder based on the retraining-related information; receiving a second RS from the base station; generating a second CFI by compressing the received second RS by the reconfigured first encoder; and sending the second CFI to the base station, wherein the retraining-related information may include at least one of information indicating retraining of the encoder, information about a discarded node of the encoder, or identifier information of the encoder, and the identifier information of the first encoder may further include the number of discards.

[0030] Beneficial effects

[0031] According to the present invention, when using an autoencoder for CSI feedback, the UE compresses the received CSI-RS and reports it to the base station, which results in less overhead than codebook-based CSI feedback. Furthermore, by appropriately determining the CSI period based on channel conditions and the UE's hardware, the autoencoder can be adaptively trained. Furthermore, by performing training and retraining of the autoencoder when discarding is applied to the encoder's nodes, the training of the autoencoder can be accelerated. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a conceptual diagram illustrating a first exemplary embodiment of a communication system.

[0033] Figure 2 is a block diagram illustrating a first exemplary embodiment of a communication node constituting a communication system.

[0034] Figure 3 is a block diagram illustrating a first exemplary embodiment of a communication node performing communications.

[0035] Figure 4a is a block diagram illustrating a first exemplary embodiment of a transmit path.

[0036] Figure 4b is a block diagram illustrating a first exemplary implementation of a receive path.

[0037] Figure 5 is a conceptual diagram illustrating a first exemplary embodiment of a system frame in a communication system.

[0038] Figure 6 is a conceptual diagram illustrating a first exemplary embodiment of a subframe in a communication system.

[0039] Figure 7 is a conceptual diagram illustrating a first exemplary embodiment of time slots in a communication system.

[0040] Figure 8 is a conceptual diagram illustrating a first exemplary embodiment of time-frequency resources in a communication system.

[0041] Figure 9 This is a conceptual diagram used to describe the offline training process of the autoencoder and the operation of the UE to obtain the autoencoder.

[0042] Figure 10 This is a conceptual diagram used to describe the structure of the encoder and decoder of an autoencoder and the channel estimation scenario using the autoencoder.

[0043] Figure 11 is a conceptual diagram of an encoder model showing application of dropout to specific nodes of the encoder.

[0044] Figure 12 is a sequence diagram illustrating operations between a base station and a UE during a CFI feedback procedure for online training.

[0045] Figure 13 is a sequence diagram for describing a process for performing online training of an autoencoder at a base station.

[0046] Figure 14 is a flowchart for describing a case where a base station determines that RSRP is saturated.

[0047] Figure 15is a sequence diagram for describing a process for terminating online training of an autoencoder at a base station.

[0048] Figure 16 It is a conceptual diagram for describing a case where the overall operation of the present invention is combined and executed. DETAILED DESCRIPTION

[0049] Since the present invention is susceptible to various modifications and may have a variety of forms, specific exemplary embodiments will be shown in the drawings and described in detail in the detailed description. However, it should be understood that it is not intended to limit the present invention to specific exemplary embodiments, but on the contrary, the present invention covers all modifications and alternative forms that fall within the spirit and scope of the present invention.

[0050] Relational terms such as first, second, etc. can be used to describe various elements, but these elements should not be limited by the terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the present invention, a first component can be named a second component, and a second component can be similarly named a first component. The term "and / or" means any one or combination of multiple related and described matters.

[0051] In the present invention, “at least one of A and B” may mean “at least one of A or B” or “at least one of a combination of one or more of A and B”. In addition, “one or more of A and B” may mean “one or more of A or B” or “one or more of a combination of one or more of A and B”.

[0052] In the present invention, "(re)transmission" may refer to "transmission", "retransmission" or "transmission and retransmission", "(re)configuration" may refer to "configuration", "reconfiguration" or "configuration and reconfiguration", "(re)connection" may refer to "connection", "reconnection" or "connection and reconnection", and "(re)access" may refer to "access", "reaccess" or "access and reaccess".

[0053] When it is mentioned that a certain component is “coupled” or “connected” to another component, it should be understood that the certain component is directly “coupled” or “connected” to the other component, or another component may be provided therebetween. Conversely, when it is mentioned that a certain component is “directly coupled” or “directly connected” to another component, it should be understood that no other component is provided therebetween.

[0054] The terms used in the present invention are only used to describe specific exemplary embodiments and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In the present invention, terms such as "including" or "having" are intended to indicate the presence of features, quantities, steps, operations, components, parts, or combinations thereof described in the specification, but it should be understood that these terms do not exclude the presence or addition of one or more features, quantities, steps, operations, components, parts, or combinations thereof.

[0055] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. Terms commonly used in dictionaries and already in dictionaries should be interpreted as having meanings that match the contextual meanings in the art. In this specification, unless explicitly defined, terms are not necessarily interpreted as having formal meanings.

[0056] Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. In describing the present invention, in order to facilitate a comprehensive understanding of the present invention, the same reference numerals refer to the same elements throughout the description of the accompanying drawings, and repeated descriptions thereof will be omitted. Operations according to the exemplary embodiments explicitly described in the present invention, as well as combinations of the exemplary embodiments, extensions of the exemplary embodiments, and / or variations of the exemplary embodiments may be performed. Some operations may be omitted, and the sequence of operations may be changed.

[0057] Even when a method (e.g., signal transmission or reception) performed at a first communication node among communication nodes is described in the exemplary embodiment, a corresponding second communication node may also perform a method (e.g., signal reception or transmission) corresponding to the method performed at the first communication node. That is, when the operation of a user equipment (UE) is described, the corresponding base station may perform an operation corresponding to the operation of the UE. Conversely, when the operation of a base station is described, the corresponding UE may perform an operation corresponding to the operation of the base station.

[0058] A base station may be referred to by various terms, such as Node B, evolved Node B, next generation node B (gNodeB), gNB, device, apparatus, node, communication node, base transceiver station (BTS), radio remote head (RRH), transmission reception point (TRP), radio unit (RU), roadside unit (RSU), radio transceiver, access point, access node, etc. A user equipment (UE) may be referred to by various terms, such as terminal, device, apparatus, node, communication node, end node, access terminal, mobile terminal, station, subscriber station, mobile station, portable subscriber station, on-board unit (OBU), etc.

[0059] In the present invention, signaling may be one of higher layer signaling, MAC signaling, and physical (PHY) signaling, or a combination of two or more. A message used for higher layer signaling may be referred to as a "higher layer message" or a "higher layer signaling message." A message used for MAC signaling may be referred to as a "MAC message" or a "MAC signaling message." A message used for PHY signaling may be referred to as a "PHY message" or a "PHY signaling message." Higher layer signaling may refer to the operation of sending and receiving system information (e.g., a master information block (MIB), a system information block (SIB)) and / or an RRC message. MAC signaling may refer to the operation of sending and receiving a MAC control element (CE). PHY signaling may refer to the operation of sending and receiving control information (e.g., downlink control information (DCI), uplink control information (UCI), or sidelink control information (SCI)).

[0060] In the present invention, "configuration of an operation (e.g., a transmission operation)" may refer to configuration information (e.g., information elements, parameters) required for the operation and / or signaling of information indicating the execution of the operation. "Configuration of information elements (e.g., parameters)" may refer to signaling of information elements. In the present invention, "signal and / or channel" may refer to a signal, a channel, or both a signal and a channel, and "signal" may be used to mean "signal and / or channel."

[0061] The communication network to which the exemplary embodiment is applied is not limited to the communication network described below, and the exemplary embodiment can be applied to various communication networks (e.g., 4G communication network, 5G communication network and / or 6G communication network). Here, "communication network" can be used interchangeably with the term "communication system".

[0062] Figure 1 is a conceptual diagram illustrating a first exemplary embodiment of a communication system.

[0063] like Figure 1 As shown, the communication system 100 may include multiple communication nodes 110-1, 110-2, 110-3, 120-1, 120-2, 130-1, 130-2, 130-3, 130-4, 130-5 and 130-6. In addition, the communication system 100 may further include: a core network (e.g., a serving gateway (S-GW), a packet data network (PDN) gateway (PDN gateway, P-GW), a mobility management entity (MME)). When the communication system 100 is a 5G communication (e.g., an NR system), the core network may include an access and mobility management function (AMF), a user plane function (UPF), a session management function (SMF), etc.

[0064] The plurality of communication nodes 110 to 130 may support communication protocols specified in the 3rd Generation Partnership Project (3GPP) standard (eg, LTE communication protocol, LTE-A communication protocol, NR communication protocol, etc.). The plurality of communication nodes 110 to 130 may support the following technologies: code division multiple access (CDMA), wideband CDMA (WCDMA), time division multiple access (TDMA), frequency division multiple access (FDMA), orthogonal frequency division multiplexing (OFDM), filtered OFDM, cyclic prefix OFDM (CP-OFDM), discrete Fourier transform spread OFDM (DFT-s-OFDM), orthogonal frequency division multiple access (OFDMA), single carrier FDMA (SC-FDMA), non-orthogonal multiple access (NOMA), generalized frequency division multiplexing (GFDM), filter bank multi-carrier (filter bank multi-carrier) The communication nodes may include a universal filtered multi-carrier (FBMC) technology, a universal filtered multi-carrier (UFMC) technology, a space division multiple access (SDMA) technology, etc. Each of the plurality of communication nodes may have the following structure.

[0065] Figure 2 is a block diagram illustrating a first exemplary embodiment of a communication node constituting a communication system.

[0066] like Figure 2As shown, the communication node 200 may include at least one processor 210, a memory 220, and a transceiver 230 connected to a network for performing communication. In addition, the communication node 200 may further include an input interface device 240, an output interface device 250, a storage device 260, etc. Each component included in the communication node 200 can communicate with each other when connected through a bus 270.

[0067] The processor 210 may execute a program stored in at least one of the memory 220 and the storage device 260. The processor 210 may refer to a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated processor on which the method according to the embodiment of the present invention is executed. Each of the memory 220 and the storage device 260 may be composed of at least one of a volatile storage medium and a non-volatile storage medium. For example, the memory 220 may include at least one of a read-only memory (ROM) and a random access memory (RAM).

[0068] Reference again Figure 1 , the communication system 100 may include a plurality of base stations 110-1, 110-2, 110-3, 120-1, and 120-2, and a plurality of terminals 130-1, 130-2, 130-3, 130-4, 130-5, and 130-6. The communication system 100 including the base stations 110-1, 110-2, 110-3, 120-1, and 120-2, and the terminals 130-1, 130-2, 130-3, 130-4, 130-5, and 130-6 may be referred to as an "access network." Each of the first base station 110-1, the second base station 110-2, and the third base station 110-3 may form a macro cell, and each of the fourth base station 120-1 and the fifth base station 120-2 may form a small cell. The fourth base station 120-1, the third terminal 130-3, and the fourth terminal 130-4 may be within the cell coverage of the first base station 110-1. In addition, the second terminal 130-2, the fourth terminal 130-4, and the fifth terminal 130-5 may be within the cell coverage of the second base station 110-2. In addition, the fifth base station 120-2, the fourth terminal 130-4, the fifth terminal 130-5, and the sixth terminal 130-6 may be within the cell coverage of the third base station 110-3. In addition, the first terminal 130-1 may be within the cell coverage of the fourth base station 120-1, and the sixth terminal 130-6 may be within the cell coverage of the fifth base station 120-2.

[0069] Here, each of the plurality of base stations 110-1, 110-2, 110-3, 120-1, and 120-2 may refer to a Node-B, an evolved Node-B (eNB), a gNB, an advanced base station (ABS), a high reliability base station (HR-BS), a base transceiver station (BTS), a radio base station, a radio transceiver, an access point, an access node, a radio access station (RAS), a mobile multihop relay-base station (MMR-BS), a relay station (RS), an advanced relay station (ARS), a high reliability relay station (HR-RS), a home NodeB (HNB), a home eNodeB (HeNB), a road side unit (RSU), a radio remote head (RRH), a transmission point (TTR), or a base station. point (TP), transmission and reception point (TRP), etc.

[0070] Each of the multiple terminals 130-1, 130-2, 130-3, 130-4, 130-5 and 130-6 can refer to user equipment (UE), terminal equipment (TE), advanced mobile station (AMS), high reliability mobile station (HR-MS), terminal, access terminal, mobile terminal, station, subscriber station, mobile station, portable subscriber station, node, device, on-board unit (OBU), etc.

[0071] On the other hand, each of the plurality of base stations 110-1, 110-2, 110-3, 120-1, and 120-2 may operate in the same frequency band or in different frequency bands. The plurality of base stations 110-1, 110-2, 110-3, 120-1, and 120-2 may be connected to each other via an ideal backhaul or a non-ideal backhaul, and exchange information with each other via the ideal or non-ideal backhaul. Furthermore, each of the plurality of base stations 110-1, 110-2, 110-3, 120-1, and 120-2 may be connected to the core network via an ideal or non-ideal backhaul. Each of the multiple base stations 110-1, 110-2, 110-3, 120-1 and 120-2 can transmit a signal received from the core network to the corresponding terminal 130-1, 130-2, 130-3, 130-4, 130-5 or 130-6, and transmit a signal received from the corresponding terminal 130-1, 130-2, 130-3, 130-4, 130-5 or 130-6 to the core network.

[0072] In addition, each of the multiple base stations 110-1, 110-2, 110-3, 120-1 and 120-2 can support multiple-input multiple-output (MIMO) transmission (e.g., single-user MIMO (SU-MIMO), multi-user MIMO (MU-MIMO), massive MIMO, etc.), coordinated multipoint (CoMP) transmission, carrier aggregation (CA) transmission, transmission in unlicensed frequency bands, sidelink communication (e.g., device-to-device (D2D) communication, proximity service (ProSe)), Internet of Things (IoT) communication, dual connectivity (DC), etc. Here, each of the multiple terminals 130-1, 130-2, 130-3, 130-4, 130-5, and 130-6 can perform operations corresponding to the following: operations of the multiple base stations 110-1, 110-2, 110-3, 120-1, and 120-2, and operations supported by the multiple base stations 110-1, 110-2, 110-3, 120-1, and 120-2. For example, the second base station 110-2 can transmit a signal to the fourth terminal 130-4 in SU-MIMO mode, and the fourth terminal 130-4 can receive a signal from the second base station 110-2 in SU-MIMO mode. Alternatively, the second base station 110-2 can transmit a signal to the fourth terminal 130-4 and the fifth terminal 130-5 in MU-MIMO mode, and the fourth terminal 130-4 and the fifth terminal 130-5 can receive a signal from the second base station 110-2 in MU-MIMO mode.

[0073] The first base station 110-1, the second base station 110-2, and the third base station 110-3 can transmit signals to the fourth terminal 130-4 in a CoMP transmission manner, and the fourth terminal 130-4 can receive signals from the first base station 110-1, the second base station 110-2, and the third base station 110-3 in a CoMP manner. In addition, each of the multiple base stations 110-1, 110-2, 110-3, 120-1, and 120-2 can exchange signals with the corresponding terminal 130-1, 130-2, 130-3, 130-4, 130-5, or 130-6 within the coverage area of ​​its cell in a CA manner. Each of base stations 110-1, 110-2, and 110-3 may control sidelink communication between the fourth terminal 130-4 and the fifth terminal 130-5, and thus the fourth terminal 130-4 and the fifth terminal 130-5 may perform sidelink communication under the control of the second base station 110-2 and the third base station 110-3, respectively.

[0074] On the other hand, a communication node that performs communication in a communication network may be configured as follows. Figure 3 The communication nodes shown may be Figure 2 A specific exemplary embodiment of a communication node is shown.

[0075] Figure 3 is a block diagram illustrating a first exemplary embodiment of a communication node performing communications.

[0076] like Figure 3 As shown, each of the first communication node 300a and the second communication node 300b can be a base station or a UE. The first communication node 300a can send a signal to the second communication node 300b. The transmission processor 311 included in the first communication node 300a can receive data (e.g., data units) from the data source 310. The transmission processor 311 can receive control information from the controller 316. The control information can include at least one of system information, RRC configuration information (e.g., information configured by RRC signaling), MAC control information (e.g., MAC CE), or PHY control information (e.g., DCI, SCI).

[0077] The transmit processor 311 may generate data symbols by performing processing operations on data (e.g., encoding operations, symbol mapping operations, etc.). The transmit processor 311 may generate control symbols by performing processing operations on control information (e.g., encoding operations, symbol mapping operations, etc.). In addition, the transmit processor 311 may generate synchronization / reference symbols for synchronization signals and / or reference signals.

[0078] The Tx MIMO processor 312 may perform spatial processing operations (e.g., precoding operations) on data symbols, control symbols, and / or synchronization / reference symbols. The output of the Tx MIMO processor 312 (e.g., a symbol stream) may be provided to a modulator (MOD) included in transceivers 313a through 313t. The modulator may generate modulation symbols by performing processing operations on the symbol stream and may generate signals by performing additional processing operations (e.g., analog-to-analog conversion, amplification, filtering, upconversion, etc.) on the modulation symbols. The signals generated by the modulators of transceivers 313a through 313t may be transmitted via antennas 314a through 314t.

[0079] The signal transmitted by the first communication node 300a can be received at the antenna 364a to the antenna 364r of the second communication node 300b. The signal received at the antenna 364a to the antenna 364r can be provided to the demodulator (DEMOD) included in the transceiver 363a to the transceiver 363r. The demodulator (DEMOD) can obtain samples by performing processing operations on the signal (e.g., filtering operations, amplification operations, down-conversion operations, digital conversion operations, etc.). The demodulator can perform additional processing operations on the samples to obtain symbols. The MIMO detector 362 can perform MIMO detection operations on the symbols. The receiving processor 361 can perform processing operations on the symbols (e.g., deinterleaving operations, decoding operations, etc.). The output of the receiving processor 361 can be provided to the data sink. 360 and controller 366. For example, data may be provided to data sink 360, and control information may be provided to controller 366.

[0080] On the other hand, the second communication node 300b can transmit a signal to the first communication node 300a. The transmit processor 368 included in the second communication node 300b can receive data (e.g., data units) from the data source 367 and perform processing operations on the data to generate data symbols. The transmit processor 368 can receive control information from the controller 366 and perform processing operations on the control information to generate control symbols. In addition, the transmit processor 368 can generate reference symbols by performing processing operations on reference signals.

[0081] The Tx MIMO processor 369 may perform spatial processing operations (e.g., precoding operations) on data symbols, control symbols, and / or reference symbols. The output of the Tx MIMO processor 369 (e.g., a symbol stream) may be provided to a modulator (MOD) included in transceivers 363a through 363t. The modulator may generate modulation symbols by performing processing operations on the symbol stream and may generate signals by performing additional processing operations (e.g., analog-to-analog conversion, amplification, filtering, and upconversion) on the modulation symbols. The signals generated by the modulators of transceivers 363a through 363t may be transmitted via antennas 364a through 364t.

[0082] The signal transmitted by the second communication node 300b can be received at antennas 314a to 314r of the first communication node 300a. The signals received at antennas 314a to 314r can be provided to a demodulator (DEMOD) included in transceivers 313a to 313r. The demodulator can obtain samples by performing processing operations (e.g., filtering, amplification, down-conversion, and digital conversion) on the signal. The demodulator can perform additional processing operations on the samples to obtain symbols. The MIMO detector 320 can perform MIMO detection operations on the symbols. The receive processor 319 can perform processing operations (e.g., deinterleaving, decoding, etc.) on the symbols. The output of the receive processor 319 can be provided to the data sink 318 and the controller 316. For example, data can be provided to the data sink 318, and control information can be provided to the controller 316.

[0083] Memory 315 and memory 365 may store data, control information, and / or program codes. Scheduler 317 may perform scheduling operations for communications. Figure 3 The processors 311, 312, 319, 361, 368, and 369 and the controllers 316 and 366 shown may be Figure 2 The processor 210 shown can be used to execute the methods described in the present invention.

[0084] Figure 4a is a block diagram illustrating a first exemplary embodiment of a transmit path, Figure 4b is a block diagram illustrating a first exemplary implementation of a receive path.

[0085] like Figure 4a and Figure 4bAs shown, a transmission path 410 may be implemented in a communication node transmitting a signal, and a reception path 420 may be implemented in a communication node receiving a signal. The transmission path 410 may include a channel coding and modulation block 411, a serial-to-parallel (S-to-P) block 512, an N-point inverse fast Fourier transform (N-point IFFT) block 413, a parallel-to-serial (P-to-S) block 414, a cyclic prefix (CP) addition block 415, and an up-converter (UC) 416. The reception path 420 may include a down-converter (DC) 421, a CP removal block 422, an S-to-P block 423, an N-point FFT block 424, a P-to-S block 425, and a channel decoding and demodulation block 426. Here, N may be a natural number.

[0086] In the transmit path 410, information bits may be input to a channel coding and modulation block 411. The channel coding and modulation block 411 may perform encoding and decoding operations (e.g., low-density parity check (LDPC) encoding and decoding operations, polar encoding and decoding operations, etc.) and modulation operations (e.g., quadrature phase shift keying (OPSK), quadrature amplitude modulation (QAM), etc.) on the information bits. The output of the channel coding and modulation block 411 may be a modulation symbol sequence.

[0087] The S-to-P block 412 may convert the frequency-domain modulation symbols into parallel symbol streams to generate N parallel symbol streams. N may be the IFFT size or the FFT size. The N-point IFFT block 413 may generate a time-domain signal by performing an IFFT operation on the N parallel symbol streams. The P-to-S block 414 may convert the output of the N-point IFFT block 413 (e.g., the parallel signal) into a serial signal to generate a serial signal.

[0088] The CP adding block 415 may insert a CP into the signal. The UC 416 may up-convert the frequency of the output of the CP adding block 415 to a radio frequency (RF) frequency. In addition, the output of the CP adding block 415 may be filtered in baseband before up-conversion.

[0089] The signal transmitted from transmit path 410 may be input to receive path 420. The operations in receive path 420 may be the inverse of those in transmit path 410. DC 421 may down-convert the frequency of the received signal to baseband frequency. CP removal block 422 may remove the CP from the signal. The output of CP removal block 422 may be a serial signal. S to P block 423 may convert the serial signal into parallel signals. N-point FFT block 424 may generate N parallel signals by performing an FFT algorithm. P to S block 425 may convert the parallel signals into a sequence of modulation symbols. Channel decoding and demodulation block 426 may perform a demodulation operation on the modulation symbols and may recover the data by performing a decoding operation on the result of the demodulation operation.

[0090] exist Figure 4a and Figure 4b In the present invention, discrete Fourier transform (DFT) and inverse DFT (IDFT) can be used instead of FFT and IFFT. Figure 4a and Figure 4b Each of the blocks (eg, components) in the embodiment may be implemented by at least one of hardware, software, or firmware. For example, Figure 4a and Figure 4b Some blocks in the may be implemented by software, and other blocks may be implemented by hardware or a combination of hardware and software. Figure 4a and Figure 4b In the , a block can be subdivided into multiple blocks, multiple blocks can be integrated into one block, some blocks can be omitted, and blocks that support other functions can be added.

[0091] Figure 5 is a conceptual diagram illustrating a first exemplary embodiment of a system frame in a communication system.

[0092] like Figure 5 As shown, time resources in a communication system can be divided on a frame basis. For example, system frames of a communication system can be configured continuously in the time domain. The length of a system frame can be 10 milliseconds (ms). The system frame number (SFN) can be set to one of #0 to #1023. In this case, 1024 system frames can be repeated in the time domain of the communication system. For example, the SFN of the system frame after system frame #1023 can be #0.

[0093] A system frame may include two half-frames. A half-frame may be 5 ms long. The half-frame at the beginning of the system frame may be referred to as "half-frame #0," and the half-frame at the end of the system frame may be referred to as "half-frame #1." A system frame may include 10 subframes. A subframe may be 1 ms long. The 10 subframes within a system frame may be referred to as subframe #0 to subframe #9.

[0094] Figure 6 is a conceptual diagram illustrating a first exemplary embodiment of a subframe in a communication system.

[0095] like Figure 6 As shown, a subframe may include n time slots, where n may be a natural number. Accordingly, a subframe may consist of one or more time slots.

[0096] Figure 7 is a conceptual diagram illustrating a first exemplary embodiment of time slots in a communication system.

[0097] like Figure 7 As shown, a time slot may include one or more symbols. For example, Figure 7 One slot shown in FIG may include 14 symbols. The length of a slot may vary according to the number of symbols included in the slot and the length of the symbol. Alternatively, the length of a slot may vary according to a parameter set (numerology).

[0098] The parameter sets applied to the physical signals and channels in the communication system may be variable. The parameter sets may be adjusted to meet various technical requirements of the communication system. In a communication system that applies OFDM waveform technology based on a cyclic prefix (CP), the parameter set may include a subcarrier spacing and a CP length (or CP type). Table 1 may illustrate a first exemplary embodiment of a method for configuring a parameter set for a CP-OFDM-based communication system. Depending on the frequency band in which the communication system operates, at least some of the parameter sets in Table 1 may be supported. In addition, the communication system may support parameter sets not listed in Table 1.

[0099] [Table 1]

[0100]

[0101] When the subcarrier spacing is 15 kHz (e.g., μ = 0), the length of the time slot may be 1 ms. In this case, one system frame may include 10 time slots. When the subcarrier spacing is 30 kHz (e.g., μ = 1), the length of the time slot may be 0.5 ms. In this case, one system frame may include 20 time slots.

[0102] When the subcarrier spacing is 60 kHz (e.g., μ = 2), the length of the time slot may be 0.25 ms. In this case, one system frame may include 40 time slots. When the subcarrier spacing is 120 kHz (e.g., μ = 3), the length of the time slot may be 0.125 ms. In this case, one system frame may include 80 time slots. When the subcarrier spacing is 240 kHz (e.g., μ = 4), the length of the time slot may be 0.0625 ms. In this case, one system frame may include 160 time slots.

[0103] Symbols can be configured as downlink (DL) symbols, flexible (FL) symbols, or uplink (UL) symbols. A slot consisting only of DL symbols may be referred to as a "DL slot," a slot consisting only of FL symbols may be referred to as an "FL slot," and a slot consisting only of UL symbols may be referred to as a "UL slot."

[0104] The slot format can be semi-statically configured by higher layer signaling (e.g., RRC signaling). Information indicating the semi-static slot format can be included in the system information, and the semi-static slot format can be configured to be cell-specific. In addition, the semi-static slot format can be further configured for each terminal by terminal-specific higher layer signaling (e.g., RRC signaling). The flexible symbols in the cell-specific slot format can be overwritten as downlink symbols or uplink symbols by terminal-specific higher layer signaling. In addition, the slot format can be dynamically indicated by physical layer signaling (e.g., a slot format indicator (SFI) included in the DCI). The semi-statically configured slot format can be overwritten by the dynamically indicated slot format. For example, the semi-statically configured flexible symbols can be overwritten by the SFI as downlink symbols or uplink symbols.

[0105] Reference signals may include Channel State Information-Reference Signal (CSI-RS), Sounding Reference Signal (SRS), Demodulation-Reference Signal (DM-RS), and Phase Tracking-Reference Signal (PT-RS). Channels may include Physical Broadcast Channel (PBCH), Physical Downlink Control Channel (PDCCH), Physical Downlink Shared Channel (PDSCH), Physical Uplink Control Channel (PUCCH), Physical Uplink Shared Channel (PUSCH), Physical Sidelink Control Channel (PSCCH), and Physical Sidelink Shared Channel (PSSCH). In the present invention, the control channel may refer to PDCCH, PUCCH or PSCCH, and the data channel may refer to PDSCH, PUSCH or PSSCH.

[0106] Figure 8 is a conceptual diagram illustrating a first exemplary embodiment of time-frequency resources in a communication system.

[0107] like Figure 8 As shown, a resource consisting of one OFDM symbol on the time axis and one subcarrier on the frequency axis can be defined as a "resource element (RE)". A resource consisting of one OFDM symbol on the time axis and K subcarriers on the frequency axis can be defined as a "resource element group (REG)". A REG can include K REs. REG can be used as a basic unit for resource allocation in the frequency domain. K can be a natural number. For example, K can be 12. N can be a natural number. Figure 7 In the time slot shown, N may be 14. N OFDM symbols may be used as a basic unit for resource allocation in the time domain.

[0108] In the present invention, RB may refer to a common RB (CRB). Alternatively, RB may refer to a physical RB (PRB) or a virtual RB (VRB). In a communication system, CRB may refer to an RB that constitutes a set of continuous RBs (e.g., a common RB grid) based on a reference frequency (e.g., point A). A carrier and / or bandwidth part may be mapped onto the common RB grid. That is, a carrier and / or bandwidth part may be configured with CRBs. The RBs or CRBs constituting the bandwidth part may be referred to as PRBs, and the CRB index may be appropriately converted to a PRB index within the bandwidth part.

[0109] Downlink data can be transmitted through PDSCH. The base station can send the configuration information of PDSCH (e.g., scheduling information) to the terminal through PDCCH. The terminal can obtain the configuration information of PDSCH by receiving PDCCH (e.g., downlink control information (DCI)). For example, the configuration information of PDSCH may include the modulation coding scheme (MCS) used for transmission / reception of PDSCH, time resource information of PDSCH, frequency resource information of PDSCH, and feedback resource information of PDSCH. PDSCH may refer to the radio resources for sending and receiving downlink data. Alternatively, PDSCH may refer to the downlink data itself. PDCCH may refer to the radio resources for sending and receiving downlink control information (e.g., DCI). Alternatively, PDCCH may refer to the downlink control information itself.

[0110] The terminal may monitor the PDCCH to receive the PDSCH transmitted from the base station. The base station may notify the terminal of configuration information for the PDCCH monitoring operation using a higher layer message (e.g., a Radio Resource Control (RRC) message). The configuration information for the PDCCH monitoring operation may include control resource set (CORESET) information and search space information.

[0111] CORESET information may include PDCCH DMRS information, PDCCH precoding information, and PDCCH opportunity information. PDCCH DMRS may be a DMRS used to demodulate the PDCCH. A PDCCH opportunity refers to an area where a PDCCH may potentially exist, which means that it is an area where DCI can be transmitted. A PDCCH opportunity may also be referred to as a PDCCH candidate. PDCCH opportunity information may include time resource information and frequency resource information for the PDCCH opportunity. In the time domain, the length of the PDCCH opportunity may be indicated in symbols. In the frequency domain, the size of the PDCCH opportunity may be indicated in RB units (e.g., PRB units or CRB units).

[0112] The search space information may include a core set identifier (ID) associated with the search space, a periodicity for PDCCH monitoring, and / or an offset for PDCCH monitoring. The periodicity and offset for PDCCH monitoring may each be indicated in units of time slots. In addition, the search space information may further include an index of a symbol at which the PDCCH monitoring operation starts.

[0113] A base station may configure a bandwidth part (BWP) for downlink communication. The BWP may be configured differently for each terminal. The base station may notify the terminal of BWP configuration information using higher-layer signaling. Higher-layer signaling may refer to the transmission of system information and / or RRC messages. The number of BWPs configured for a single terminal may be one or more. The terminal may receive BWP configuration information from the base station and identify the configured BWP based on the received configuration information. When multiple BWPs are configured for downlink communication, the base station may activate one or more of the multiple BWPs. The base station may transmit configuration information of the activated BWP to the terminal using at least one of higher-layer signaling, a Medium Access Control (MAC) Control Element (CE), or DCI. The base station may perform downlink communication using the activated BWP. The terminal may identify the activated BWP by receiving the configuration information from the base station and perform downlink reception on the activated BWP.

[0114] On the other hand, the current 5G NR communication system uses a codebook-based channel state information (CSI) feedback scheme to estimate downlink CSI. The codebook-based CSI feedback scheme may refer to a scheme in which the base station periodically or aperiodically sends a downlink CSI reference signal (CSI-Reference Signal, CSI-RS) to the UE, and the UE sends CSI to the base station. In this case, the CSI may include a channel quality indicator (CQI), a precoding matrix indicator (PMI), and a rank indication (RI). The base station can estimate the CSI using a predefined codebook based on the CQI, PMI, and RI reported by the UE.

[0115] Future 6G communication systems are expected to utilize unlicensed millimeter wave or terahertz frequency bands to enable ultra-high capacity, ultra-wide bandwidth, and ultra-low latency communication services. These systems are expected to utilize frequency bands several dozen times wider than the FR2 band used in 5G NR communication systems. Consequently, a CSI feedback process with higher accuracy and higher overhead may be required.

[0116] 3GPP is also discussing artificial intelligence / machine learning (AI / ML) solutions for the physical layer as a study item (SI) in Release 18. AI / ML solutions can involve training a model using output data corresponding to input data to extract specific patterns or parameters, and predicting output data for arbitrary input data based on the trained model. 3GPP has discussed the application of AI / ML solutions in technical areas such as CSI feedback, beam management, and enhanced positioning accuracy.

[0117] 3GPP SI related to AI / ML is being explored in the context of CSI feedback, beam management, and enhanced positioning accuracy. For CSI feedback using AI / ML, discussions have begun on priorities associated with reduced overhead, improved channel estimation accuracy, and CSI prediction compared to existing 5G NR. Based on discussions at the Rel-18 RAN1#109-e meeting on AI / ML for CSI feedback enhancement, it was agreed to configure the AI / ML network as a bilateral model consisting of a CSI feedback component and a CSI reconstruction component to support AI / ML-based CSI feedback. The UE and base station perform channel estimation using the CSI feedback component and the CSI reconstruction component, respectively.

[0118] In addition, various techniques have been proposed in the 3GPP Release 18 discussions to support AI / ML-based CSI feedback, including a reduction in overhead compared to the current Type 2 codebook-based CSI feedback defined in the 3GPP technical specifications, the need for AI / ML-based CSI feedback with higher accuracy than currently provided by the 3GPP technical specifications, and the determination of the quantization level of the AI / ML model.

[0119] However, in the current agreed or proposed solutions in 3GPP Release 18, no specific methods or detailed procedures for improving AI / ML performance are discussed.

[0120] Therefore, to support AI / ML-based CSI feedback solutions, which replace the codebook-based CSI feedback solutions currently used in 3GPP technical specifications, which utilize CQI, PMI, RI, etc., it is necessary to define new CSI feedback procedures and related parameters that further consider the characteristics of AI / ML beyond the currently discussed protocols or recommendations. The following description will address these aspects.

[0121] According to the agreement made in 3GPP Release 18, the AI / ML solution discussed for CSI feedback may include a model training phase and a model inference phase. The model training phase may be a phase in which the AI ​​model and / or ML model is trained using training data as input to extract parameters, etc. The model inference phase may be a phase in which inference is performed using the input data for inference based on the parameters obtained in the model training phase of the AI ​​model and / or ML model. Due to its computationally intensive nature, the model training phase typically requires a large number of iterative operations and may therefore be performed over a long period of time. In addition, the performance of the model inference may be determined by the accuracy of the parameters trained in the model training phase.

[0122] In the case of channel estimation, the model training phase of the AI ​​model and / or ML model can be interpreted as a phase of performing training using input data to obtain parameters for channel estimation, and the model inference phase can be interpreted as a phase of performing channel estimation in an actual channel environment using the parameters obtained in the model training phase of the AI ​​model and / or ML model.

[0123] 3GPP has proposed online and offline training solutions as AI / ML model training processes for AI / ML support. Online training refers to a solution that generates new training data in real time and uses it to train AI / ML models. Offline training refers to a solution that uses pre-collected data to train AI / ML models.

[0124] When online training is applied to channel estimation technology, real-time channel information is used in online training, so that accurate parameters suitable for the current channel between the base station and the UE can be obtained. Therefore, more accurate channel training is possible compared to offline training between the base station and the UE. However, AI / ML model training usually takes a considerable amount of time, and the AI / ML model training time may become much longer than the channel coherence time. Therefore, channel estimation may not be performed in a short time through online training schemes.

[0125] When offline training is applied to channel estimation technology, the base station or server performing the training can pre-train the AI / ML model using previously collected data, allowing channel estimation to be performed within the channel coherence time. However, because offline training applies parameters trained using previously collected data to the AI / ML model, accurate channel estimation may not be possible.

[0126] As mentioned above, to address the issues associated with channel estimation and feedback methods using AI / ML technologies, a combined online and offline training approach can be considered as a solution. For example, a process can be considered in which an AI / ML model is trained through offline training to derive parameters that enable a certain level of accurate channel estimation. Therefore, offline training can be a rough training process. Subsequently, to account for real-time channels, a fine-tuning process can be considered using the online trained AI / ML model.

[0127] However, for AI models and / or ML models, a channel estimation process and scheme for jointly performing offline training and online training has not yet been agreed upon in 3GPP. In addition, when performing channel estimation technology by combining offline training and online training for AI models and / or ML models, the parameters obtained through offline training may overfit to previously collected data, and even if online training is subsequently performed in real time, parameters that meet a certain level of accurate channel estimation for the AI ​​model and / or ML model may not be obtained.

[0128] Therefore, the present invention provides a CSI feedback process that takes into account the characteristics of the offline trained AI / ML model when performing channel estimation using an AI / ML scheme. Specifically, a scenario is considered in which a base station (or server) performs offline training of the AI / ML model using collected data and then performs online training of the AI / ML model to ensure a certain level of accurate channel estimation. In addition, a channel estimation technology will be described, which includes a process for determining the period and dimension of the feedback signal reported by the UE to the base station during the CSI feedback process, and a retraining process based on online training using discarding to solve the overfitting problem.

[0129] The present invention considers scenarios where downlink channels are estimated using AI / ML models. However, the present invention is not limited to downlink channel estimation. The present invention can also be applied to scenarios where uplink channels are estimated using the same or similar schemes as those described in this invention. However, for ease of description, the following describes the case of downlink channel estimation.

[0130] In the present invention, parameters for AI / ML-based CSI feedback according to the present invention are defined to replace at least one of the CQI, PMI, and RI reported by the UE to the base station in the codebook-based CSI feedback process. A process in which the UE reports the parameters defined in the present invention to the base station to perform online training of the AI / ML model will be described. For example, in order to use the AI / ML model for CSI feedback, the base station and the UE can use an autoencoder. In this case, an environment is considered in which the base station has a decoder for the autoencoder and the UE has an encoder for the autoencoder. This is based on the assumption that the base station has a decoder and the UE has an encoder for the downlink CSI feedback process. However, if uplink channel estimation is also considered, each of the base station and the UE can have both an encoder and a decoder.

[0131] In the present invention, an autoencoder may refer to a network that compresses input data into a signal with a small dimension and then restores the compressed data back to its original data form. The autoencoder may include an encoder that compresses the data and a decoder that restores the compressed data.

[0132] Next, a case where the autoencoder according to the present invention is used for downlink channel estimation will be described below.

[0133] The base station can send CSI-RS to the UE through a wireless channel. Correspondingly, the UE can receive CSI-RS from the base station through the wireless channel. Since the CSI-RS received by the UE is received through the wireless channel, the CSI-RS received by the UE may be degraded (or distorted) due to the influence of the wireless channel. The CSI-RS received by the UE through the wireless channel will be referred to as "received CSI-RS". The UE can use an encoder to compress the received CSI-RS. The UE can send the compressed received CSI-RS to the base station. If necessary, the UE can also send the compressed received CSI-RS and additional information to the base station. The base station can receive the compressed received CSI-RS from the UE or can receive the compressed received CSI-RS and additional information. The base station can restore the received CSI-RS by decoding the compressed received CSI-RS via a decoder. Since the base station knows the CSI-RS that has been transmitted, the base station can use the restored received CSI-RS and the CSI-RS sent by the base station to estimate the downlink channel. The base station can use the estimated channel information to train the autoencoder.

[0134] According to the method proposed in this invention, during the training process of the autoencoder, the UE only needs to generate and report the compressed received CSI-RS through the encoder during the channel estimation process. Therefore, compared with the codebook-based CSI feedback according to the current 5G NR technical specifications, the UE uplink overhead according to this invention can be significantly reduced.

[0135] In the present invention, the autoencoder can be trained through offline training using pre-collected data. Offline training can be performed at the base station or a specific server. In the present invention, for ease of description, it will be assumed that the offline training of the autoencoder is performed at the base station. However, offline training of the autoencoder can also be performed at the server. If offline training of the autoencoder is performed at the server, the server can pre-provide the trained autoencoder to the base station. Alternatively, the UE can receive the offline trained autoencoder from the base station or server.

[0136] Figure 9 This is a conceptual diagram used to describe the offline training process of the autoencoder and the operation of the UE to obtain the autoencoder.

[0137] like Figure 9 As shown, a database 930, a base station 910 and a UE 920 are shown. Figure 9 In order to simplify the drawing, the base station 910 is shown to have an autoencoder 911 and to perform offline training of the autoencoder 911. However, the server ( Figure 9 ) instead of the base station 910, the base station 910 may perform offline training of the autoencoder and provide the offline trained autoencoder 911 to the base station 910.

[0138] Database 930 can store collected data for offline training of autoencoder 911. In the present invention, since autoencoder 911 is described as being used for downlink channel estimation, the collected data can be channel information. For example, the channel information can be obtained through simulation. As another example, the channel information can include various measurements actually obtained between multiple base stations and multiple UEs and collected from the base stations. The collected data stored in database 930 can be preprocessed data in a form suitable for training the autoencoder according to the present invention.

[0139] The base station 910 may include an autoencoder 911 according to the present invention. As described above, the autoencoder 911 may include an encoder 9111 and a decoder 9112. For ease of description, the decoder of the autoencoder will be referred to as the decoder 9112, and the encoder of the autoencoder will be referred to as the encoder 9111. The base station 910 may receive collected data from the database 930 and may perform offline training of the autoencoder 911. Thus, the base station 910 may utilize the collected data to perform offline training of the autoencoder. The autoencoder may be configured to perform a compression and recovery scheme.

[0140] The process of training the encoder 9111 and the decoder 9112 through offline training will be briefly described. The encoder 9111 can encode the collected data received from the database 930. The encoded data can be provided to the decoder 9112 (S941). Therefore, the decoder 9112 can receive the compressed data from the encoder 9111 (S941).

[0141] The decoder 9112 can decode and restore the compressed data. The decoder 9112 can use the restored information to obtain channel information. Based on the obtained channel information, the decoder 9112 can generate control information for adjusting the encoder 9111 and the decoder 9112. The control information for adjusting the encoder 9111 and the decoder 9112 will be described in more detail with reference to the drawings described later. The decoder 9112 can send the control information to the encoder 9111 (S942). Therefore, the encoder 9111 can receive the control information from the decoder 9112 (S942).

[0142] The encoder 9111 can be trained by applying the received control information. The encoder 9111 can re-encode the collected data based on the coding scheme applied with the control information. The subsequent process can be iterated in the same manner as described above. Such iterations can be performed until the decoded information at the decoder 9112 meets the pre-configured level.

[0143] In other words, the encoded data generated by the encoding operation of the encoder 9111 is provided to the decoder 9112, and the decoder 9112 can check whether the channel estimation is performed correctly by decoding the encoded data. If the channel estimation is not performed correctly, in other words, if the channel estimation is not performed within the expected error range, the encoder 9111 and / or the decoder 9112 can generate control information for controlling the operation of the encoder 9111 and the decoder 9112, and the control information can be applied to the encoder 9111 and / or the decoder 9112. The encoder 9111 and / or the decoder 9112, to which the control information for controlling the operation of the encoder 9111 and the decoder 9112 is applied, can iteratively perform an offline training process through steps S941 and S942, and the error is reduced.

[0144] When offline training of the encoder 9111 and the decoder 9112 is completed, the base station 910 may transmit the autoencoder 911 to be used for channel estimation or the encoder 9111 constituting the autoencoder 911 to the UE 920 . Figure 9 FIG. 9 shows a case where only the encoder 9111 constituting the autoencoder is transmitted to the UE 920. Figure 9 In the figure, since the encoder 9111 in the UE 920 is the same as the encoder 9111 in the base station 910, the same figure mark is used.

[0145] It should be noted that Figure 9 Only the autoencoder 911 that the base station 910 and the UE 920 may store or possess is shown. Therefore, the base station 910 may include the above reference Figure 2 If the offline training of the autoencoder 911 is performed at a specific server and the base station 910 receives the autoencoder 911 from the server, then in addition to Figure 2 In addition to the components shown in , the base station 910 may further include an interface for communicating with a server. In addition, the memory 220 of the base station 910 may store the autoencoder 911. The processor 210 of the base station 910 may serve as a main body for executing and / or controlling the operation of the autoencoder 911 described below.

[0146] As described above, the UE 920 may store the offline-trained autoencoder 911 or the offline-trained encoder 9111 of the autoencoder 911 received from the base station. As another example, the UE 920 may pre-store the offline-trained encoder 9111. As another example, the UE 920 may receive the offline-trained autoencoder 911 or the offline-trained encoder 9111 of the autoencoder 911 from the server.

[0147] If the UE 920 pre-stores the encoder 9111 trained offline, the UE 920 may update the encoder 9111 based on the version information or information about the offline training date of the encoder 9111 trained offline. For example, if the UE 920 pre-stores the encoder 9111 trained offline, the UE 920 may receive the version information or information about the offline training date of the encoder 911 that the base station 910 intends to use from the base station 910 in advance.

[0148] If the version information of the self-encoder 911 received from the base station 910 is the same as the version information of the encoder 9111 stored in the UE 920, the UE 920 can use the stored encoder 9111 without modification. On the other hand, if the version information of the self-encoder 911 received from the base station 910 is different from the version information of the encoder 9111 stored in the UE 920 (even if the version of the self-encoder 911 of the base station 910 is lower), the UE 920 can receive and update the encoder 9111 from the base station 910.

[0149] It should be noted that in Figure 9 , only the encoder 9111 of the self-encoder 911 is used to illustrate the configuration of the UE 920. The UE 920 may include the above reference Figure 2 The UE 920 may further include an interface and / or sensor for user convenience. The encoder 9111 of the autoencoder 911 may be stored in the memory 220 of the UE 920. The processor 210 of the UE 920 may perform control of the operation of the autoencoder 911 described in the present invention.

[0150] As above Figure 9 As described above, the process of training the autoencoder 911 (in other words, the encoder 9111 and the decoder 9112) through offline training can be called "first training", "pre-training" or "offline training".

[0151] The base station 910 may perform offline training of the encoder 911 using previously collected data, and the offline trained encoder 9111 may be delivered to the UE 920. In this manner, first training, pre-training, or offline training may be performed through coarse training for channel estimation.

[0152] Next, the encoder 9111 and the decoder 9112 constituting the autoencoder 911 will be described.

[0153] Figure 10 This is a conceptual diagram used to describe the structure of the encoder and decoder of an autoencoder and the channel estimation scenario using the autoencoder.

[0154] like Figure 10 , an encoder 9111 of the autoencoder 911, a wireless channel 1001, and a decoder 9112 of the autoencoder 911 are shown. Therefore, the encoder 9111 and the decoder 9112 may be an encoder and a decoder, respectively, for which offline training has been performed.

[0155] according to Figure 9, the encoder 9111 may be maintained by the UE 920, and the decoder 9112 may be maintained by the base station 910. This is because the present invention assumes the case where the encoder 911 is used for downlink channel estimation. In the case of uplink channel estimation, the encoder 9111 may be maintained by the base station 910, and the decoder 9112 may be maintained by the UE 920. When performing both downlink and uplink channel estimation, the base station 910 and the UE 920 may each maintain both the encoder 9111 and the decoder 9112 trained offline.

[0156] refer to Figure 10 , the encoder 9111 may include an input layer 1010, a hidden layer 1020, and an output layer 1030. Figure 10 , for ease of description, the case where hidden layer 1020 is configured as a single layer is shown. However, hidden layer 1020 may be composed of two or more layers. Generally, as the number of layers in hidden layer 1020 increases, the processing time of encoder 9111 becomes longer, and more accurate channel estimation can be performed. Conversely, as the number of layers in hidden layer 1020 decreases, the processing time of encoder 9111 becomes shorter, and the accuracy of channel estimation may decrease.

[0157] The configuration of the encoder 9111 will be described. The input layer 1010 of the encoder 9111 may be composed of a plurality of input nodes. Figure 10 In FIG. 1 , for the convenience of description, the input layer 1010 of the encoder 9111 is shown to be composed of six nodes. Figure 10 Each of the nodes of the input layer 1010 of the encoder 9111 can be called a neuron. In other words, Figure 10 Each black dot in the image can correspond to a node or a neuron.

[0158] Each node in the input layer 1010 can receive input data. Here, the input data can be referred to as an input variable. The input variable provided to the input layer 1010 can be information used for channel estimation. For example, the input variable can be the CSI-RS received as described above. Each node constituting the input layer 1010 can be connected to a node in the hidden layer 1020. When the hidden layer 1020 has two or more layers, each node in the input layer 1010 can be connected to each node in the first hidden layer. When each node in the input layer 1010 is connected to a node in the hidden layer 1020, the connection between the nodes can be represented by a trained weight.

[0159] The configuration of the hidden layer 1020 will be described. As described above, for ease of description, Figure 10The hidden layer 1020 is shown as a single layer. However, the hidden layer 1020 may be composed of multiple layers. If the hidden layer 1020 is composed of multiple layers, each of the hidden layers may also be composed of multiple nodes. For example, if the hidden layer 1020 is composed of two layers, each node of the second hidden layer may be connected to each node of the first hidden layer. Each node of the second hidden layer may also be connected to each node of the output layer 1030.

[0160] In the following, for the convenience of description, the following description will be made of Figure 10 The case of a single hidden layer 1020 is shown.

[0161] As described above, each node of the hidden layer 1020 can be connected to an input node of the input layer 1010. Figure 10 In the example of FIG, the hidden layer 1020 is shown to be composed of four nodes. The number of nodes in the hidden layer 1020 is not limited to Figure 10 In other words, the hidden layer 1020 may consist of five or more nodes, or three or fewer nodes.

[0162] Each node constituting the hidden layer 1020 may also be connected to a node of the output layer 1030. Based on this configuration, each node of the hidden layer 1020 may provide a connection between a node of the input layer 1010 and a node of the output layer 1030. In this case, when an input variable is transmitted from a node of the input layer 1010 to a node of the output layer 1030, each node of the hidden layer 1020 may transmit information calculated based on a weighted sum obtained through offline and / or online training.

[0163] Next, the output layer 1030 of the encoder 9111 will be described.

[0164] The output layer 1030 of the encoder 9111 may output a latent variable based on information received from a corresponding node of the hidden layer 1020. Here, the latent variable may refer to a value output from a node existing in the latent space of the output layer 1030 of the encoder 9111. Figure 10 In the example of , since the number of nodes included in the latent space is 2, the latent space can be represented as M, and the value of M can be set to 2. Since the number of latent space nodes in the output layer 1030 (that is, the value of M) is smaller than the number of input variable nodes in the input layer 1010, the encoding operation of the encoder 9111 according to the present invention can be understood as a compression operation.

[0165] Using the configuration of the encoder 9111 described above, the operation of generating information for CSI feedback by the UE 920 can be described as follows.

[0166] The base station 910 may transmit an RS for channel estimation (e.g., a CSI-RS or a CSI-RS used for online training of an autoencoder) to the UE 920. Therefore, the UE 920 may receive the CSI-RS transmitted by the base station 910. In this case, since the CSI-RS is transmitted via the wireless channel 1001, it may be degraded or distorted by the wireless channel 1001. Accordingly, the UE 920 may compress the received CSI-RS using the encoder 9111.

[0167] The encoder 9111 can transmit the CSI-RS received at the input layer 1010 to the hidden layer 1020 and the output layer 1030 based on the information of the offline training. In this case, the received CSI-RS can be calculated using the weighted sum obtained by offline training, and can be sequentially compressed and output through the corresponding nodes of the input layer 1010, the hidden layer 1020, and the output layer 1030. In other words, the encoder 9111 can compress the received CSI-RS based on offline training. The compressed CSI-RS can be a latent variable value, which is the output data of the corresponding node of the output layer 1030. The UE 920 can generate a channel feature indicator (CFI) using the latent variable value. The CFI according to the present invention can be expressed as the following equation 1.

[0168] [Equation 1]

[0169]

[0170] In Equation 1, N can be determined as the product of the number of nodes M in the latent space and the latent variable dimension (LVD). In the case of encoder 9111, the number of nodes M in the latent space can be determined as the number of output nodes in output layer 1030. LVD will be described in more detail below.

[0171] The UE 920 may transmit the CFI to the base station 910 through the wireless channel 1001. Accordingly, the base station 910 may receive the CFI transmitted by the UE 920 through the wireless channel 1001. The UE 920 may transmit the CFI to the base station 910 when online training of the autoencoder is required and / or when channel estimation is performed using the autoencoder.

[0172] Base station 910 may periodically transmit CSI-RS. Furthermore, base station 910 may transmit additional CSI-RS when needed. Accordingly, UE 920 may periodically receive CSI-RS. UE 920 may also receive additional CSI-RS. When base station 910 requires channel estimation, UE 920 may transmit CFI to base station 910 instead of transmitting CSI feedback.

[0173] Furthermore, when online training of the autoencoder 911 is required, the base station 910 may notify the UE 920 of the CFI reporting period, and the UE 920 may report the CFI to the base station 910 based on the CFI reporting period. In this case, the base station 910 may transmit the CSI-RS based on the CFI reporting period, or may transmit the CSI-RS at a predetermined period. The CFI reporting period will be described in more detail below.

[0174] The base station 910 can use the CFI received from the UE 920 as an input to the decoder 9112 to recover the received CSI-RS from the CFI. As described above, the received CSI-RS may be degraded or distorted by the wireless channel 1001 after being transmitted by the base station 910. Therefore, the CSI-RS recovered by the base station 910 using the CFI received from the UE 920 as an input to the decoder 9112 can theoretically be the same information as the CSI-RS received by the UE 920.

[0175] The operation of the decoder 9112 may correspond to the reverse process of the encoding process of the encoder 9111. In other words, the decoder 9112 may include an input layer 1040, a hidden layer 1050, and an output layer 1060. The input layer 1040 of the decoder 9112 may be composed of the same number of nodes as the output layer 1030 of the encoder 9111 in order to process CFI. Accordingly, the nodes of the input layer 1040 of the decoder 9112 may form an input latent space and may be composed of two nodes that are the same as the latent space of the output layer 1030 of the encoder 9111. The hidden layer 1050 of the decoder 9112 may be composed of the same number of nodes as the hidden layer 1020 of the encoder 9111. According to Figure 10 In the example of , the hidden layer 1050 of the decoder 9112 can be composed of four nodes. Similarly, the output layer 1060 of the decoder 9112 can be composed of the same six nodes as the input layer 1010 of the encoder 9111.

[0176] like Figure 10 As shown, the base station 910 can recover the received CSI-RS from the CFI reported by the UE 920. The base station 910 can obtain a reference signal received power (RSRP) value from the recovered received CSI-RS.

[0177] When performing online training, the base station 910 may receive two or more CFIs from the UE 920. The base station 910 may obtain an RSRP value from each of the received CFIs. The base station may compare the obtained RSRP values ​​to determine whether the RSRP values ​​are saturated. Here, saturation may refer to a state in which the RSRP value does not change by more than a certain level within a specific time or a specific number of times.

[0178] For example, assuming that the number of times is three, the RSRP value becomes saturated as follows.

[0179] The base station 910 may determine whether two or more RSRP values ​​have been acquired before the current time t at which the current RSRP value is acquired. It may be assumed that the time at which the current RSRP value is acquired is denoted as t, and the times at which the previous RSRP values ​​are acquired are denoted as t-1 and t-2, respectively. If the RSRP(t-2) acquired at time t-2, the RSRP(t-1) acquired at time t-1, and the RSRP(t) acquired at time t are equal or within a predetermined range, the base station 910 may determine that the RSRP value is saturated. For example, if the predetermined range is 2 dBm, the RSRP(t-2) value is -48.4 dBm, the RSRP(t-1) value is -49 dBm, and the RSRP(t) value is -48.7 dBm, the base station 910 may determine that the RSRP value is saturated. On the other hand, if the RSRP(t-2) value is -44.2 dBm, the RSRP(t-1) value is -47.4 dBm, and the RSRP(t) value is -48.7 dBm, the base station 910 may determine that the RSRP values ​​are not saturated.

[0180] If the RSRP value is saturated, the base station 910 can complete online training and determine whether the saturated RSRP value meets the expected level of RSRP. Here, the expected level of RSRP can be determined based on the distance between the base station 910 and the UE 920, the channel environment between them, etc.

[0181] There may be various methods for the base station 910 to obtain the expected RSRP value during the online training of the autoencoder. For example, the base station 910 may receive the location information of the UE 920 and determine the expected RSRP value based on the distance between the base station 910 and the UE 920. As another example, the UE 920 may provide CSI feedback according to the 3GPP technical specification during or before the online training of the autoencoder. Accordingly, the base station 910 may determine the expected RSRP value based on the CSI feedback from the UE 920. As another method, the base station 910 may determine (or estimate) the expected RSRP value based on the transmission power level of the signal for reporting CFI by the UE 920 and the power level of the received signal containing the CFI. In this case, the UE 920 may send information about the transmission power level for reporting CFI as additional information.

[0182] Typically, the training saturation of an AI / ML model can be assessed based on how closely the input data reconstructed from the output of the trained network matches the original input. In other words, saturation can be determined based on the accuracy of data recovery.

[0183] However, when using Figure 10 When the structure shown performs channel estimation in an actual wireless communication environment, the actual channel environment corresponding to the actual input data is unknown. Therefore, it is impossible to determine saturation based on recovery accuracy. Accordingly, in the wireless communication environment described in the present invention, since the goal is to maximize communication quality, the training saturation of the AI ​​model or ML model can be determined based on whether the RSRP value is saturated.

[0184] Therefore, as reference Figure 10 and Figure 11 As described above, when the saturated RSRP value is within the expected RSRP range, the base station can complete the training of the encoder 9111 and the decoder 9112, and use the parameters obtained through online training for channel estimation in the model inference stage.

[0185] On the other hand, if the saturated RSRP value does not meet the desired RSRP value (or range), the base station 910 may determine that the parameters trained through offline training are overfitting to the previously collected data. In this case, to resolve the issue, the base station 910 may instruct the UE 920 to apply discarding to specific nodes of the encoder 9111.

[0186] When UE 920 instructs base station 910 to apply discarding to a specific node of encoder 9111, UE 920 can perform discarding on the corresponding node and retrain the model through online training. In the present invention, discarding can refer to a technique for partially omitting a neural network to solve the overfitting problem. In the present invention, discarding can be implemented by assigning a discarding probability to each node that is subject to discarding. The following describes the case of applying discarding to encoder nodes with reference to the accompanying drawings.

[0187] Figure 11 is a conceptual diagram of an encoder model showing application of dropout to specific nodes of the encoder.

[0188] Figure 11 The encoder 9111 shown may be an encoder stored in the UE 920. In other words, the encoder 9111 may be an encoder received from the server or base station 910, as described above with reference to Figure 9 However, Figure 11 Shows the case where a particular node has been dropped.

[0189] Figure 11 The input layer 1110 of the encoder 9111 shown has Figure 10 The encoder 9111 has the same number of nodes as the input layer 1010, Figure 11 The hidden layer 1120 of the encoder 9111 shown has Figure 10 The encoder 9111 has the same number of nodes as the hidden layer 1020, and Figure 11 The output layer 1130 of the encoder 9111 shown has Figure 10 The output layer 1030 of the encoder 9111 has the same number of nodes.

[0190] and Figure 10 compared to, Figure 11 1 shows a case where the second node 1121 of the hidden layer 1120 has been discarded. Accordingly, Figure 11 The connections between the input layer 1110 and the hidden layer 1120 in the encoder 9111 shown can be Figure 10 The connections between the input layer 1010 and the hidden layer 1020 in the encoder 9111 are shown to be different. In other words, Figure 11 Each node of the input layer 1110 shown may be in a state of being disconnected from the second node 1121 of the hidden layer 1120 .

[0191] and Figure 10 In comparison, due to Figure 11 The second node 1121 in the hidden layer 1120 is dropped, and the connection between the hidden layer 1120 and the output layer 1130 can also be Figure 10In other words, Figure 11 The second node 1121 of the hidden layer 1120 is shown to be in a state of not being connected to any node of the output layer 1130. In other words, Figure 11 The second node 1121 of the hidden layer 1120 represented by a white dot in FIG may be in a state of being disconnected from both the input layer 1110 and the output layer 1130 according to a discard instruction.

[0192] The connections between the remaining nodes in the hidden layer 1120 except the second node 1121 are the same as Figure 10 . In the present invention, specific nodes can be discarded by setting their weights to zero, thereby ensuring that the corresponding nodes are discarded. In other words, since the weight of the connection from the input layer 1110 to the second node 1121 of the hidden layer 1120 is zero, no value is transmitted. In addition, the second node 1121 of the hidden layer 1120 can also apply a weight of zero to the nodes of the output layer 1130, thereby not transmitting any value.

[0193] In the following description of the present invention, the processes and parameters required to specifically perform the above operations will be defined. Based on these definitions, the processes for performing the operations shown below will be described. The processes presented in the present invention may include the following.

[0194] (1) Process for determining the transmission period of CSI feedback parameters (i.e., CFI)

[0195] (2) The process used to determine the latent variable dimension (LVD) of the latent space

[0196] (3) The base station sends CSI-RS to the UE, and the UE generates CFI based on the received CSI-RS and reports the CFI to the base station.

[0197] (4) The base station uses CFI to train the decoder part of the base station and determines the time to terminate the training through the saturation threshold

[0198] (5) The process of evaluating training performance by RSRP threshold after training is terminated

[0199] (6) a process for retraining the UE's encoder by dropping specific nodes when the training performance is unsatisfactory, and

[0200] (7) A process for terminating the training of the autoencoder when the training performance is satisfactory.

[0201] The above process (1) to process (7) will be described in more detail below. The process shown above can be used to train an autoencoder and can represent an operation of performing fine-tuning on the downlink channel estimate on an autoencoder that has undergone offline training through online training.

[0202] [A]. CFI feedback process for online training

[0203] The following describes the CFI feedback process for online training. A UE that has received an offline-trained autoencoder from a base station or a server can perform the CFI feedback process. In this case, the base station may have an entire offline-trained autoencoder or a decoder of an offline-trained autoencoder. The UE may have an entire offline-trained autoencoder or an encoder of an offline-trained autoencoder. Since offline training of the autoencoder has been described above, repeated descriptions will be omitted.

[0204] In the following description, channel estimation will be described by assuming downlink channel estimation. However, the same or similar method can be applied to uplink channel estimation. In the present invention, for downlink estimation, the base station can send CSI-RS to the UE, and the UE can generate CFI by compressing the received CSI-RS. The UE can perform online training by sending the generated CFI to the base station through the CSI feedback process. In the following description, the training of the autoencoder can refer to Figure 9 , the training of the encoder 9111, the training of the decoder 9112, or the training of both the encoder 9111 and the decoder 9112. Even when the terms "training" and "learning" are used interchangeably, they may refer to a process of performing training of an autoencoder, an encoder, or a decoder through online training.

[0205] Figure 12 is a sequence diagram illustrating operations between a base station and a UE during a CFI feedback procedure for online training.

[0206] In reference Figure 12 Previously, based on Figure 9 UE 920 and base station 910 are described in the same manner as in Figure 9 The same reference numerals as used in the reference numerals are used for the UE 920 and the base station 910. In addition, the encoder 9111 of the self-encoder may have Figure 10 The configuration of the encoder described in , and the decoder 9112 of the self-encoder can also have Figure 10 Therefore, the encoder 9111 and the decoder 9112 can be in an offline training state.

[0207] At step S1200, UE 920 and base station 910 may coordinate online training of an autoencoder. Coordination of online training of an autoencoder may be initiated by base station 910 instructing UE 920 to perform online training of the autoencoder, or by UE 920 requesting the base station to perform online training of the autoencoder. For online training of an autoencoder, base station 910 and UE 920 must have the same autoencoder. If UE 920 does not have the autoencoder required for online training, base station 910 may provide UE 920 with the entire autoencoder required for online training or the encoder of the autoencoder to be trained at UE 920.

[0208] As described below, there may be two or more autoencoders with different network sizes. Each of the autoencoders may be selected according to circumstances, and online training may be performed on at least one autoencoder.

[0209] At step S1200, base station 910 may also provide UE 920 with information required for autoencoder training. For example, base station 910 may provide UE 920 with table information and / or a start time for online training, as described below. UE 920 may also provide its location information to base station 910, either voluntarily or upon request from base station 910.

[0210] If necessary, the base station 910 may transmit a reference signal (e.g., CSI-RS) to the UE 920. The UE 920 may report an RSRP value, which is information about the received power of the CSI-RS transmitted by the base station 910, to the base station 910. The process in which the base station 910 transmits the CSI-RS to the UE 920 and the UE 920 reports the CSI to the base station 910 may be performed based on the CSI feedback process defined in the current 3GPP technical specifications.

[0211] Although the coordination process for online training of the autoencoder at step S1200 may require additional processes besides those described above, descriptions of all processes and operations will be omitted.

[0212] In step S1210, base station 910 may determine a CFI transmission period. In the present invention, the CFI transmission period may be determined based on the network size and / or channel coherence time of autoencoder 911. In addition to the network size and channel coherence time of the autoencoder described in the present invention, additional factors may be considered when determining the CFI transmission period. However, this article will only discuss these two factors.

[0213] a. Determine the CFI transmission period based on the network size of the autoencoder

[0214] According to an exemplary embodiment of the present invention, the base station 910 may determine the CFI transmission period based on the network size of the autoencoder 911. Here, the network of the autoencoder 911 may refer to the network of the autoencoder 911 trained by the base station 910 or the server through offline training. The network size NS of the autoencoder 911 may be determined as the sum of the width and height of the network constituting the encoder 9111 or the decoder 9112 of the autoencoder 911.

[0215] Here, the network width can be defined as the Figure 10 The number of nodes of a layer in the encoder 9111 is shown. Since the decoder 9112 has a configuration that performs the opposite operation of the encoder 9111, the network width can also be defined as the number of nodes forming a layer in the decoder 9112.

[0216] The network height may refer to the number of layers that make up the encoder 9111. Figure 10 In the example of FIG, the encoder 9111 is shown as having a single hidden layer 1020. However, the hidden layer 1020 can be composed of two or more layers. Therefore, the network height can vary depending on the number of hidden layers 1020.

[0217] Based on the above description, the network width can be understood as the total number of nodes that constitute the input layer, hidden layer, and output layer.

[0218] exist Figure 10 In the encoder 9111 shown, the input layer 1010, the hidden layer 1020, and the output layer 1030 are each composed of one layer. In the example shown in the encoder 9111, the number of nodes constituting the input layer 1010 is 6, the number of nodes constituting the hidden layer 1020 is 4, and the number of nodes constituting the output layer 1030 is 2. Therefore, Figure 10 The network size NS of the encoder 9111 shown can be interpreted as "6+4+2=12".

[0219] As described above, the number of autoencoders trained offline may be multiple. Each of the autoencoders trained offline may have a different network size. Each of the autoencoders trained offline may be maintained by both the base station 910 and the UE 920. The CFI transmission period corresponding to the network size NS of each autoencoder having a different size may be preconfigured to differ depending on the network size.

[0220] Consider a case where there are three autoencoders with different sizes. Each of the autoencoders can be referred to as a first autoencoder with a first network size (NS#1), a second autoencoder with a second network size (NS#2), and a third autoencoder with a third network size (NS#3). When the network sizes are set such that NS#1 < NS#2 < NS#3, the CFI transmission periods can be preconfigured to different values as shown in Table 2 below.

[0221] [Table 2]

[0222] Network size CFI transmission period NS#1 CFI_P#1 NS#2 CFI_P#2 NS#3 CFI_P#3

[0223] Each of the CFI transmission periods (CFI_P#1, CFI_P#2, and CFI_P#3) shown in Table 2 can be a period with a different time interval. In the case of the above network sizes, each of the CFI transmission periods (CFI_P#1, CFI_P#2, and CFI_P#3) can be set such that CFI_P#1 < CFI_P#2 < CFI_P#3.

[0224] When determining the network size of the autoencoder, the base station 910 can consider various factors. For example, the base station 910 can determine the network size of the autoencoder based on the available power of the UE 920 and / or the channel environment. Here, determining the network size of the autoencoder can be understood as selecting an autoencoder. In other words, when the base station 910 selects the first autoencoder, it selects the network size (NS#1) corresponding to the first autoencoder; when the base station 910 selects the second autoencoder, it selects the network size (NS#2) corresponding to the second autoencoder; when the base station 910 selects the third autoencoder, it selects the network size (NS#3) corresponding to the third autoencoder.

[0225] First, the case where the base station 910 determines the network size of the autoencoder based on the available power of the UE 920 is described below.

[0226] The base station 910 can determine the network size of the autoencoder used for online training with the UE 920 based on information about available power pre-reported by the UE 920. When the available power pre-reported by the UE 920 is low, that is, when the remaining battery power of the UE 920 is equal to or less than a preset threshold, the base station 910 can select an autoencoder with a small network size. Conversely, when the remaining battery power of the UE 920 is equal to or greater than the preset threshold, the base station 910 can select an autoencoder with a large network size. A large network size of the autoencoder means that the UE 920 needs to perform calculations on many layers and a large number of nodes when performing encoding. When the autoencoder has a large network size, the encoding time becomes longer due to the need for the UE 920 to perform calculations on many layers and a large number of nodes when performing encoding. Therefore, when the autoencoder has a large network size, the power consumption of the UE 920 increases, and the encoding time also becomes longer. In other words, the time required for the UE 920 to obtain the CFI becomes longer. Therefore, the base station 910 can determine the network size of the autoencoder based on the available power of the UE 920. Once the network size of the autoencoder is determined, the CFI period can be determined as shown in Table 2. When online learning of an autoencoder with a small network size is determined, the CFI period becomes shorter, and when online learning of an autoencoder with a large network size is determined, the CFI period can become longer.

[0227] Next, a case where the base station 910 determines the network size of the autoencoder based on the channel environment between the UE 920 and the base station 910 is described below.

[0228] In the previous step S1200, the base station 910 may have previously identified the rate of change of the channel environment between the UE 920 and the base station 910. The base station 910 can determine the autoencoder for performing online learning based on the rate of change of the channel environment. For example, when the rate of change of the channel environment is high, the base station 910 can select an autoencoder with a small network size. On the contrary, when the rate of change of the channel environment is low or the channel environment hardly changes, the base station 910 can select an autoencoder with a large network size. As described above, when the network size is large, the encoding time of the UE 920 becomes longer. Therefore, when the channel changes rapidly, the base station 910 can select an autoencoder with a small network size to cope with the rapid channel changes. On the contrary, when the channel changes are small or the channel hardly changes, the base station 910 can select an autoencoder with a large network size.

[0229] In the above description, a method of selecting an autoencoder using each of the available power of the UE 920 and the rate of change of the channel environment between the UE 920 and the base station 910 has been described. However, it is also possible to select an autoencoder by simultaneously considering both the available power of the UE 920 and the rate of change of the channel environment between the UE 920 and the base station 910.

[0230] The reason for mapping the network size of the self-encoder and the CFI transmission period in the present invention is described below.

[0231] Base station 910 may receive CFI from UE 920 during the online training process and input the received CFI into the decoder of the autoencoder for decoding. In this case, if the CFI reporting period is too short, UE 920 may fail to report CFI using the compressed CSI-RS. In addition, even when CFI is reported from UE 920 to base station 910, if the CFI is reported before the decoding process at the decoder of base station 910 is completed, the received CFI may not be correctly processed.

[0232] Conversely, if the CFI reporting cycle is too long, the online training itself may be delayed.

[0233] On the other hand, for ease of description, Table 2 only illustrates three network sizes and three CFI transmission periods. However, the network size and CFI period according to the present invention are not limited to three, and each of the network size and the CFI period corresponding to the network size can be four or more.

[0234] When using the mapping table in Table 2, the base station 910 may pre-provide the table information in Table 2 to the UE 920. In other words, at step S1200, the base station 910 may pre-provide the table information in Table 2 to the UE 920.

[0235] When the base station 910 pre-provides mapping information such as Table 2 to the UE 920, the base station 910 may transmit the information to the UE 920 through various signaling. For example, the mapping information such as Table 2 transmitted to the UE 920 may be transmitted using a system information block (SIB) or may be included in a radio resource control (RRC) reconfiguration message or a medium access control (MAC)-control element (CE) message.

[0236] As another example, mapping information such as Table 2 may be transmitted to the UE 920 through downlink control information (DCI).

[0237] As another example, mapping information such as that shown in Table 2 may be transmitted to the UE 920 using the second message (message 2, Msg2) of the four-step RACH procedure or message B (message B, MsgB) of the two-step RACH procedure.

[0238] As yet another example, if an RRC signaling message is newly defined to transmit mapping information such as that in Table 2, the mapping information may also be transmitted to the UE 920 using the newly defined RRC signaling message.

[0239] b. Determine the CFI transmission period based on the channel coherence time

[0240] Hereinafter, a method of determining a CFI transmission period based on a channel coherence time will be described.

[0241] The channel coherence time may refer to a duration during which the channel gain does not change. The channel coherence time is typically due to the mobility of the UE, but may also be affected by the surrounding channel environment in addition to the mobility of the UE. As the channel coherence time becomes shorter, the UE 920 may need to report CFI to the base station 910 more frequently. In addition, when the channel coherence time is shorter, the effective duration for the UE 920 to report CFI becomes shorter, and therefore, the number of times CFI is reported also decreases. As described above, when the number of CFI reports decreases, the number of CFIs that can be used for online training also decreases, and sufficient training will not be performed. When the online training of the autoencoder is insufficient, the autoencoder may obtain parameters with lower accuracy, and thus accurate channel estimation may not be possible.

[0242] Therefore, in the present invention, the CFI reporting period may be determined based on the channel coherence time (CT), and an example of the CFI reporting period based on the channel coherence time (CT) is shown in Table 3.

[0243] [Table 3]

[0244] Channel coherence time CFI transmission period CT#1 CFI_P#1 CT#2 CFI_P#2 CT#3 CFI_P#3

[0245] In Table 3, the first channel coherence time (CT#1), the second channel coherence time (CT#2), and the third channel coherence time (CT#3) are assumed such that CT#1 has the shortest time value, CT#2 has a longer time value than CT#1, and CT#3 has the longest time value. The opposite is also possible. For ease of description, the following description assumes that the time values ​​are set such that CT#1 < CT#2 < CT#3.

[0246] In the case where the relationship between the channel coherence time values is CT#1 < CT#2 < CT#3, the CFI transmission periods (CFI_P#1, CFI_P#2, CFI_P#3) can be set to increase in the order of CFI_P#1 < CFI_P#2 < CFI_P#3. In other words, for the first coherence time (CT#1), the CFI transmission period (CFI_P#1) is the shortest, and for the third coherence time (CT#3), the CFI transmission period (CFI_P#3) is the longest. This is because as the channel coherence time becomes shorter, the CFI transmission period can be set shorter to obtain more inputs for the decoder of the base station 910.

[0247] On the other hand, Table 3 classifies the channel coherence time into three levels for description, but in actual implementation, only two levels can be configured, or four or more levels can be configured. The number of levels of the channel coherence time can be preset, or can be dynamically determined by the base station 910 according to the channel environment.

[0248] When using the mapping table of Table 3, that is, when using the CFI transmission period mapped to the channel coherence time, the base station 910 can pre-provide the table information of Table 3 to the UE 920. In other words, in step S1200, the base station 910 can pre-provide the table information of Table 3 to the UE 920.

[0249] When the base station 910 pre-provides mapping information such as Table 3 to the UE 920, the base station 910 can send this information to the UE 920 through various signaling schemes. For example, the mapping information such as Table 3 sent to the UE 920 can be transmitted using SIB, or can be transmitted by being included in the RRC reconfiguration message or the MAC-CE message.

[0250] As another example, the mapping information such as Table 3 can be sent to the UE 920 through DCI.

[0251] As another example, the mapping information such as Table 3 can be sent to the UE 920 using the second message (Message 2, Msg2) of the four-step RACH procedure or the Message B (Message B, MsgB) of the two-step RACH procedure.

[0252] As yet another example, if a new RRC signaling message is defined for transmitting the mapping information such as Table 3, the mapping information can also be sent to the UE 920 using the newly defined RRC signaling message.

[0253] c. Determine the CFI transmission period based on the autoencoder size and channel coherence time

[0254] In the above part a of executing step S1210 , the operation of determining the CFI transmission period based on the size of the autoencoder has been described, and in the above part b, the operation of determining the CFI transmission period based on the channel coherence time has been described.

[0255] As described above, these can be used individually, or only one of them can be used. In other words, when determining the network size, only the size of the autoencoder can be considered, or only the channel coherence time can be considered. In addition, the CFI transmission period can be determined by combining the method of utilizing the network size (Table 2) and the method of utilizing the channel coherence time (Table 3). For example, the base station 910 can determine the CFI transmission period by considering both the network size described in Table 2 and the channel coherence time described in Table 3. In this case, a new table can be defined taking into account the contents of Tables 2 and 3. Since the new table taking into account Tables 2 and 3 can take various forms, specific examples are not provided in the present invention.

[0256] As another method, the base station 910 may determine a CFI transmission period by considering a channel coherence time, and then determine a network size of an autoencoder between the base station 910 and the UE 920 based on the determined CFI period, as described in Table 2.

[0257] In addition, when determining the CFI period as described in Tables 2 and 3 above, the base station 910 may pre-request information from the UE 920 at step S1200 and may receive the corresponding information from the UE 920. The information requested by the base station 910 from the UE 920 may be information about the available power of the UE 920, as described in Table 2 above. To obtain such information, the base station 910 may request the UE 920 via a UE Information Request message. After receiving the UE Information Request message, the UE 920 may generate a UE Information Response message as a response and may return the requested information to the base station 910.

[0258] As another example, when information such as the power available to the UE 920 at a specific time is required, for example, a maximum transmission power value or a remaining battery level at the corresponding time, the base station 910 may request such information from the UE 920 and may receive the corresponding information from the UE 920. In this case, the UE 920 may transmit the information requested by the base station 910 using one of uplink control information (UCI), the first message (Msg1) in the four-step RACH procedure, or message A (MsgA) in the two-step RACH procedure.

[0259] As another example, if a new RRC signaling message is defined for the UE 920 to transmit information requested by the base station 910, the newly defined RRC signaling message may be used.

[0260] Through one or a combination of two or more of the above methods, the base station 910 can determine the CFI transmission period in step S1210.

[0261] In step S1220, the base station 910 can determine the latent variable dimension (LVD) of the autoencoder for performing online learning. The LVD of the autoencoder can be determined based on the channel resolution or based on the latency requirement. When determining the LVD of the autoencoder, in addition to the channel resolution and latency requirement described in the present invention, other factors can also be considered. However, in the present invention, only the case where the LVD determination considers the above two factors will be described.

[0262] a. Determine LVD based on channel resolution

[0263] In the present invention, the channel resolution can be the number of downlink channels that the base station 910 aims to distinguish. For example, when the base station 910 forms downlink channels using multiple beamformings, the channel resolution can be determined based on the number of downlink beams. Accordingly, as the base station forms narrower beams or aims to perform more accurate channel estimation, the channel resolution can be higher. The LVD can represent the dimension required to represent the latent variable values in the latent space. The LVD can also be interpreted as the number of bits required to represent a single latent variable value. Therefore, the higher the LVD, the more accurately the latent variable can be represented. Accordingly, the mapping between the channel resolution (CR) and the LVD can be exemplified as in Table 4 below.

[0264] [Table 4]

[0265] Channel resolution Latent variable dimension (LVD) CR#1 LVD#1 CR#2 LVD#2 CR#3 LVD#3

[0266] In Table 4, it is assumed that the channel resolution (CR) is such that the first channel resolution (CR#1) is the lowest, the second channel resolution (CR#2) is higher than the first channel resolution (CR#1) and lower than the third channel resolution (CR#3), and the third channel resolution (CR#3) is the highest. Then, the relationship between the channel resolutions can be CR#1 < CR#2 < CR#3.

[0267] As described above, the higher the channel resolution, the more bits can be in the LVD. Therefore, according to the relationship between the channel resolutions, the LVD values in Table 4 can also have the relationship of LVD#1 < LVD#2 < LVD#3.

[0268] Although Table 4 exemplifies the case of using three levels of channel resolution, this is only an exemplary embodiment. Only two channel resolution values can be used, or four or more channel resolution values can be used. The number of channel resolutions according to the present invention can be preset, or can be adaptively changed by the base station 910 according to the channel environment.

[0269] In addition, the mapping information between the channel resolution and the LVD illustrated in Table 4 may be pre-transmitted by the base station 910 to the UE 920. For example, in step S1200, the base station 910 may pre-transmit the mapping information between the channel resolution and the LVD illustrated in Table 4 to the UE 920. In this case, the base station 910 may transmit the mapping information between the channel resolution and the LVD to the UE 920 through various signaling schemes.

[0270] For example, mapping information such as Table 4 transmitted to the UE 920 may be transmitted using the SIB, or may be transmitted by being included in an RRC reconfiguration message or a MAC-CE message.

[0271] As another example, mapping information such as Table 4 may be transmitted to the UE 920 through DCI.

[0272] As another example, mapping information such as that shown in Table 4 may be transmitted to the UE 920 using the second message (message 2, Msg2) of the four-step RACH procedure or message B (message B, MsgB) of the two-step RACH procedure.

[0273] As another example, if a new RRC signaling message is defined to transmit mapping information such as that in Table 4, the mapping information may also be sent to the UE 920 using the newly defined RRC signaling message.

[0274] b. Determine LVD based on latency requirements

[0275] Hereinafter, a case where the LVD is determined based on the delay requirement will be described.

[0276] As described above, LVD refers to the dimension required to represent the potential variable values ​​in the latent space, and can be interpreted as the number of bits required to represent the potential variable values. As LVD increases, the amount of CFI (in bits) reported by UE 920 to base station 910 through the CSI feedback process increases. As the amount of CFI increases, the time required for online training of the autoencoder may increase. As described above, when the time required for online training increases, the delay requirement requested by UE 920 may not be met. Here, the delay requirement may be a requirement derived from the characteristics of UE 920 itself, or a requirement derived from the type of service provided to UE 920. For example, when UE 920 is a specific node in a time-sensitive network, there may be a delay requirement in the time-sensitive network. In this case, UE 920 may have a delay requirement based on the characteristics of the node in the time-sensitive network.

[0277] As another example, when providing Ultra Reliable Low Latency Communication (URLLC) services to UE 920, UE 920 may have latency requirements according to the characteristics of the corresponding services. Thus, the latency requirements may exist based on the characteristics of the network in which UE 920 operates (or the characteristics of the UE) and the latency requirements from the service itself.

[0278] In the present invention, the LVD may be determined based on such latency requirements, and an example of the mapping between the latency requirements and the LVD may be as shown in Table 5 below.

[0279] [Table 5]

[0280] Latency requirements Latent variable dimension (LVD) LR#1 LVD#1 LR#2 LVD#2 LR#3 LVD#3

[0281] In Table 5, it is assumed that the latency requirements are such that LR#1 corresponds to the shortest latency, LR#2 corresponds to a latency longer than LR#1, and LR#3 corresponds to the longest latency. The relationship between the latency requirements can thus be LR#1 < LR#2 < LR#3. The LVD values corresponding to the respective latency requirements may also have the relationship LVD#1 < LVD#2 < LVD#3. In other words, the shorter the latency requirement, the smaller the LVD can be, while the longer the latency requirement, the larger the LVD can be.

[0282] Although Table 5 exemplifies the case of using three latency requirements, this is only an exemplary implementation, and the configuration may include only two latency requirements or four or more latency requirements. The number of such latency requirements and channel resolutions may be preset, or may be adaptively changed by the base station 910 according to the channel environment.

[0283] In addition, the mapping information between the latency requirements and the LVD exemplified in Table 5 may be pre-transmitted by the base station 910 to UE 920. For example, in step S1200, the base station 910 may pre-transmit the mapping information between the latency requirements and the LVD exemplified in Table 5 to UE 920. In this case, the base station 910 may send the mapping information between the latency requirements and the LVD to UE 920 through various signaling schemes.

[0284] For example, the mapping information such as Table 5 sent to UE 920 may be transmitted using SIB, or may be transmitted by being included in the RRC reconfiguration message or the MAC-CE message.

[0285] As another example, the mapping information such as Table 5 may be sent to UE 920 through DCI.

[0286] As another example, mapping information such as that shown in Table 5 may be transmitted to the UE 920 using the second message (message 2, Msg2) of the four-step RACH procedure or message B (message B, MsgB) of the two-step RACH procedure.

[0287] As another example, if a new RRC signaling message is defined to transmit mapping information such as that in Table 5, the mapping information may also be sent to the UE 920 using the newly defined RRC signaling message.

[0288] c. Determine LVD based on channel resolution and latency requirements

[0289] In the above part a of executing step S1220 , the operation of determining the LVD based on the channel resolution has been described, and in the above part b, the operation of determining the LVD based on the delay requirement has been described.

[0290] As mentioned above, these can be used individually or only one of them can be used. In other words, when determining the LVD, only the channel resolution can be considered, or only the latency requirement can be considered. In addition, the LVD can be determined by considering both the case of using the channel resolution (Table 4) and the case of using the latency requirement (Table 5).

[0291] In addition, when determining the LVD as described in Tables 4 and 5, information about the UE 920 may be required. Therefore, in this case, the base station 910 may request the UE 920 for the information required for the LVD determination at step S1200 as described above, and may receive the information required for the LVD determination from the UE 920.

[0292] Various types of signaling may be used as a method for the UE 920 to provide UE information to the base station 910. For example, the UE 920 may use one of UCI or UE assistance information.

[0293] As another example, when the UE 920 is capable of supporting training of the autoencoder, the UE 920 may pre-report information to the base station 910 using the first message (Msg1) in the four-step RACH procedure or message A (MsgA) in the two-step RACH procedure.

[0294] As another example, if a new RRC signaling message is defined to transmit information requested by the base station 910, the UE 920 may provide UE information using the newly defined RRC signaling message.

[0295] As another example, the base station 910 may request UE information from the UE 920 and receive corresponding information from the UE 920. In this case, the base station 910 may send a UE information request message to the UE 920, and the UE 920 may include the requested information in a UE information response message and report it to the base station 910.

[0296] At step S1230 , the base station 910 may send the CFI transmission period information determined at step S1210 and the potential variable dimension information determined at step S1220 to the UE 920 .

[0297] An example of the information transmitted from the base station 910 to the UE 920 at step S1230 may be shown in Table 6 below.

[0298] [Table 6]

[0299]

[0300] As shown in the example of Table 6, the base station 910 may transmit a CFI transmission period to the UE 920. The CFI transmission period may be determined based solely on the network size of the autoencoder as described in Table 2, may be determined based solely on the channel coherence time as described in Table 3, or may be determined by considering both the network size of the autoencoder and the channel coherence time.

[0301] In addition, as shown in Table 6, the base station 910 may send LVD information to the UE 920. The LVD may be determined based only on the channel resolution as described in Table 4, may be determined based only on the latency requirement as described in Table 5, or may be determined by considering both the channel resolution and the latency requirement.

[0302] In this exemplary embodiment of the present invention, it is assumed that base station 910 determines the CFI transmission period by considering the size of the autoencoder and the channel coherence time. Furthermore, in this exemplary embodiment of the present invention, it is assumed that base station 910 determines the LVD by considering the channel resolution and latency requirements.

[0303] According to another exemplary embodiment of the present invention, in addition to the above factors, the base station 910 can determine the CFI transmission period and LVD by further considering various factors such as frequency resource usage, the number of UEs connected to the base station 910, the RSRP value of the UE, and the available power of the UE.

[0304] For example, when the number of UEs attempting to communicate with the base station 910 is large or frequency resource usage is high, the base station 910 may already be in a state where a large amount of frequency resources are used for communication. Accordingly, the amount of frequency resources available at the base station 910 may be insufficient. In this case, in order to provide interference-free communication services across all frequency bands or between UEs while the autoencoder is being trained, the base station 910 may set a long CFI transmission period and a low LVD.

[0305] As another example, the lower the RSRP of the UE attempting to communicate with the base station 910 , the base station 910 may determine a shorter CFI transmission period and a higher LVD in order to improve the reliability of the CFI signal.

[0306] As another example, the base station 910 may determine a network size based on available power reported by each UE, and may determine a CFI transmission period and a LVD based on the determined network size.

[0307] At step S1230, various types of signaling may be used when the base station 910 transmits the CFI transmission period and LVD information to the UE 920. Such signaling schemes may include the signaling schemes described in Tables 2 to 5 above.

[0308] In step S1230 , the UE 920 may receive the CFI transmission period and LVD information from the base station 910 .

[0309] At step S1240, the UE 920 may determine an encoder to be trained based on the CFI transmission period and LVD information received from the base station 910. As described above at step S1200, the UE 920 may receive two or more autoencoders (or encoders of autoencoders) trained offline from the base station 910 or the server. In this case, the UE 920 may have received autoencoders having two or more network sizes. Therefore, based on the CFI transmission period information and LVD information received from the base station 910 at step S1230, the UE 920 may determine an autoencoder to be used for online training.

[0310] To briefly describe it again, the CFI transmission period can be determined based on the network size and the channel coherence time, as described above. Therefore, the UE 920 can identify the network size of the autoencoder based on the CFI transmission period information. In addition, when determining the LVD information, it can be determined Figure 10 The number of nodes in the output layer 1030 of the encoder 9111 is shown in . Based on the above information, the UE 920 can select an autoencoder for online training.

[0311] At step S1250, the base station 910 may transmit a CSI-RS for online training (or learning) of the autoencoder to the UE 920. In this case, the CSI-RS may be a CSI-RS independently defined for online training of the autoencoder, or may be a CSI-RS transmitted periodically. Therefore, at step S1250, the UE 920 may receive the CSI-RS from the base station 910.

[0312] In step S1260, the UE 920 may use the received CSI-RS as input data for the encoder 9111 of the self-encoder 911 for online training. The CSI-RS received at the UE 920 may be the same as that in the above Figure 10 . In other words, since the received CSI-RS has passed through wireless channel 1001, it may have been attenuated and distorted compared to the CSI-RS transmitted by base station 910. UE 920 can generate CFI by using the received CSI-RS as input to encoder 9111. CFI can be determined according to Equation 1 described above and can also be expressed as Equation 2 shown below.

[0313] [Equation 2]

[0314] CFI = (number of nodes in the latent space) × (latent variable dimension)

[0315] According to Equation 2, CFI can be generated based on the number of nodes M in the latent space of the autoencoder and the LVD information indicated in step S1230. In this case, the number of nodes M in the latent space can be determined based on the determination of the autoencoder. In other words, when the network size is determined, the number of nodes M in the latent space of the corresponding autoencoder can be determined. Figure 10 In the example of , the number of nodes M in the latent space is 2, and the LVD is represented by the number of bits representing each node forming the latent space. Figure 10 In this case, the CFI as expressed in Equation 2 can be expressed as the product of 2 and LVD.

[0316] At step S1270, the UE 920 may transmit CFI to the base station 910 based on the CFI transmission period. In this case, the CFI transmission period may be the CFI transmission period transmitted by the base station 910 to the UE 920 at step S1230. Therefore, at step S1270, the base station 910 may receive CFI from the UE 920 based on the CFI transmission period. At step S1270, the base station 910 may receive CFI from the UE 920.

[0317] In this case, the CFI transmitted by the UE 920 to the base station 910 may be reported in various forms. For example, since the CFI is a report on the CSI-RS, the CFI may be transmitted to the base station 910 through a measurement report message of a CSI reporting procedure, which is a CSI-RS feedback procedure according to the current 3GPP technical specifications.

[0318] As another method, if a new RRC signaling message is defined for reporting CFI, the CFI may be transmitted to the base station 910 through the newly defined RRC signaling.

[0319] The CFI reported by the UE 920 to the base station 910 at step S1270 is compared with the scheme defined in the current 3GPP technical specifications.

[0320] According to the current 3GPP technical specifications, the UE needs to generate CSI expressed as RI, CQI, and PMI values ​​based on the CSI-RS received from the base station, and needs to perform the CSI feedback process based on the type 1 codebook or the type 2 codebook. On the other hand, the method according to the present invention can generate CFI by utilizing an encoder that can express the received CSI-RS itself in a lower dimension, and the CFI can be provided to the base station as feedback. Therefore, the CFI feedback process according to the present invention can be defined as a new feedback process for AI / ML-based CSI feedback.

[0321] According to the CFI transmission method in which the UE 920 transmits the CFI to the base station 910 at step S1270, a method for mapping latent variables to CFI may be provided. The mapping relationship between the latent variables and the CFI may be determined in various ways.

[0322] [B] The process of training and retraining the autoencoder through online training

[0323] In the present invention described below, a process for training and retraining an autoencoder by online training during channel estimation using AI / ML will be described. The autoencoder used for online training can be in a state of being pre-trained by offline training. In addition, the base station can pre-have the entire autoencoder or part of the autoencoder (for example, in the case of downlink channel estimation, the base station can pre-have a decoder). The UE can also pre-have the entire autoencoder or part of the autoencoder (for example, in the case of downlink channel estimation, the UE can also pre-have an encoder). Reference has been made to Figure 9 The case where the base station and / or UE has the entire autoencoder or a part thereof and the configuration thereof are described, so the same description will be omitted.

[0324] When the base station and the UE have the autoencoder as described above, the base station and the UE can perform fine-tuning of the autoencoder through online training. The process for acquiring information for the fine-tuning process has been described in Section [A].

[0325] In section [B] according to an exemplary embodiment of the present invention, a process for training an autoencoder by online training for fine-tuning at a base station using the acquired information and a process for retraining the autoencoder will be described. Therefore, the operation described in section [B] may be an operation based on the information acquired in section [A]. However, the operation described in section [B] may also use information acquired through the CSI reporting process according to the current 3GPP technical specifications. In the following description, the process for training and retraining the autoencoder by online training will be described based on the description of section [A].

[0326] Figure 13 is a sequence diagram for describing a process for performing online training of an autoencoder at a base station.

[0327] In the description Figure 13 Previously, it was assumed that the base station and UE correspond to the reference Figures 9 to 12 Therefore, it should be noted that the reference numerals of UE 920 and base station 910 are the same as those of UE 920 and base station 910. Figure 9 The same reference numerals are used as described in Figure 13 In the description of Figure 12 process.

[0328] In step S1300, the base station 910 may receive CFI from the UE 920 and may input the received CFI to the decoder of the self-encoder owned by the base station 910. The received CFI may be the same as that in the above Figure 12 The RSRP value corresponds to the value reported by UE 920 to base station 910 during the process of step S1270. At step S1300, base station 910 may decode the received CFI using a decoder and obtain the CSI-RS received by UE 920. Therefore, base station 910 may obtain an RSRP value from the received CSI-RS. Base station 910 may determine whether the obtained RSRP value is in a saturated state. Here, the RSRP saturation state may refer to a phenomenon in which the RSRP value does not increase despite the base station continuously performing online training using the decoder.

[0329] To determine the RSRP saturation state, base station 910 may use information such as Table 7 below.

[0330] [Table 7]

[0331]

[0332] In Table 7, the condition for determining whether to continue training can use the saturation count (COUNTSAT) and the maximum saturation count (MAXCOUNT_SAT). If the saturation count is less than the maximum saturation count, training or learning can be continued. If the saturation count is equal to the maximum saturation count, it can be determined as saturated and training or learning can be stopped. Now refer to Figure 14 This is described in more detail.

[0333] Figure 14 is a flowchart for describing a case where a base station determines that RSRP is saturated.

[0334] At step S1400, the base station 910 may set a saturation count (COUNTSAT) value to 0. At step S1402, as described above, the base station 910 may restore the received CSI-RS by inputting the received CFI into a decoder maintained by the base station 910. Therefore, the base station 910 may obtain an RSRP value using the restored received CSI-RS.

[0335] In step S1404, the base station 910 may compare the acquired RSRP value with the previous RSRP value. If the received CFI is the first received CFI, the acquired RSRP value may correspond to the initial RSRP value. Therefore, if there is no previous RSRP value (i.e., the first CFI value is acquired) or if the current RSRP value is not equal to the previous RSRP value, the base station 910 may return to step S1400.

[0336] If the RSRP value is not obtained for the first time after receiving the CFI, there is a previous RSRP value. Therefore, the base station 910 can compare the RSRP value obtained by decoding with the RSRP value obtained during the previous training, and in step S1404, can check whether the two values ​​are equal.

[0337] If the two values ​​are determined to be equal in step S1404, the base station 910 may proceed to step S1406 and increment the saturation count value by 1. On the other hand, if the two values ​​are determined to be unequal in step S1406, the base station 910 may proceed to step S1400. In this case, to prevent an infinite loop when the RSRP values ​​are unequal and the process repeatedly proceeds to step S1400, a repeat count limit value may be set. Since the method of using a repeat count limit value to prevent an infinite loop is well known, further description will be omitted.

[0338] In the above description, it has been assumed that the two RSRP values ​​are equal. However, if the two values ​​exist within a specific range, they can be considered equal.

[0339] After incrementing the saturation count value by 1 in step S1606, the base station 910 may proceed to step S1408. In step S1408, the base station 910 may determine whether the saturation count value is equal to the maximum saturation count value (MAXCOUNT_SAT). If the saturation count value is equal to the maximum saturation count value, the base station 910 may determine that RSRP has reached saturation and may terminate the process. Figure 14 routine.

[0340] On the other hand, if the saturation count value is not equal to the maximum count value, the base station 910 may return to step S1402 and repeat the training according to the present invention.

[0341] Here, the maximum saturation count value may be actively set by the base station 910 based on various factors such as the channel environment between the base station 910 and the UE 920 or the mobility of the UE 920. For example, in an environment where the channel changes rapidly or the mobility of the UE 920 is high, the base station 910 may set the maximum count value to a smaller value to perform training at a faster speed. Conversely, in a situation where the channel changes minimally and the mobility of the UE 920 is low, the base station 910 may set the maximum count value to a larger value for more accurate training.

[0342] Reference again Figure 13 In step S1300, the base station 910 may determine whether the RSRP has reached saturation based on the above description. If the RSRP is determined to be saturated, the base station may proceed to step S1510.

[0343] On the other hand, despite Figure 13 This example shows a case where base station 910 determines RSRP saturation based on CFI reported by UE 920. UE 920 can determine whether RSRP has reached saturation. In this case, if the same CFI value is generated each time CFI is generated based on received CSI-RS, UE 920 can determine that RSRP is saturated. As another example, UE 920 can directly measure the RSRP of the received CSI-RS and determine whether RSRP is saturated.

[0344] In the case where the UE 920 determines that the RSRP is saturated, the base station 910 may provide information on a maximum saturation count value (MAXCOUNT_SAT) to the UE 920. In this case, the maximum saturation count value information may be provided to the UE 920 through various forms of signaling.

[0345] For example, the base station 910 may provide the maximum saturation count value information to the UE 920 through an SIB, a DCI, a MAC-CE, a second message (Msg2) of a four-step RACH procedure, a message B (MsgB) of a two-step RACH procedure, or an RRC reconfiguration message. As another example, if a new RRC signaling message is defined for delivering information related to the maximum saturation count value (MAXCOUNT_SAT) to the UE 920, the base station 910 may use the new RRC signaling message to send the information related to the maximum count value to the UE 920.

[0346] If the UE 920 determines whether the RSRP has reached saturation, the UE 920 may notify the base station 910 when the RSRP is in a saturated state. In this case, various forms of signaling may be used as a method for the UE 920 to notify the base station 910 of the RSRP saturation state.

[0347] For example, the UE 920 may inform the base station 910 of the RSRP saturation state by using UCI or UE assistance information.

[0348] As another example, the UE 920 may inform the base station 910 of the RSRP saturation state by utilizing the first message (Msg1) in the four-step RACH procedure or the message A (MsgA) in the two-step RACH procedure.

[0349] As another example, if a new RRC signaling message is defined for notifying the RSRP saturation state, the UE 920 may notify the base station 910 of the RSRP saturation state using the newly defined RRC signaling message.

[0350] As yet another example, if the base station 910 requests RSRP saturation status information from the UE 920 through the UE Information Request message, the UE 920 may inform the base station 910 of the RSRP saturation status using a UE Information Response message.

[0351] The above has described a method for determining the RSRP saturation state by the base station 910 or the UE 920. In the following description, for convenience of description, it is assumed that the base station 910 determines whether the RSRP has reached saturation.

[0352] If the RSRP is in a saturated state as a result of performing step S1300 , the base station 910 may proceed to step S1310 .

[0353] In step S1310, the base station 910 may determine whether the RSRP value meets the expected level. As a method for determining whether the RSRP meets the expected level, the saturated RSRP value may be compared with the RSRP threshold RSRP. th Here, RSRP thThe parameter may indicate a certain level of RSRP value set by the base station 910 or the UE 920, and may be set based on various criteria such as the desired accuracy of channel estimation by the base station 910 and / or the UE 920 or hardware issues of the UE 920. In this case, in order to determine the accuracy of the channel estimation, etc., the UE 920 may Figure 12 The above-mentioned step S1200 reports the RSRP or CSI directly measured based on the CSI-RS received from the base station 910. Therefore, the base station 910 can determine the RSRP threshold based on the RSRP or CSI reported in step S1200.

[0354] For example, when base station 910 and / or UE 920 desire highly accurate channel estimation, base station 910 and / or UE 920 may perform retraining to achieve the desired performance by setting a higher RSRP threshold. As another example, when frequent training is not feasible due to hardware issues with UE 920, remaining battery capacity of UE 920, or channel congestion, base station 910 and / or UE 920 may set a lower RSRP threshold to reduce the number of retraining operations. The training results based on the RSRP threshold may be shown in Table 8 below.

[0355] [Table 8]

[0356] <![CDATA[Condition (RSRP threshold: RSRP th )]]> Training Results <![CDATA[RSRP≥RSRP th ]]> Terminate training <![CDATA[RSRP<RSRP th ]]> Need to retrain

[0357] According to Table 8 above, when the RSRP value is greater than or equal to the RSRP threshold RSRP th On the other hand, when the RSRP value is lower than the RSRP threshold RSRP th , the base station 910 may determine that retraining is required.

[0358] Figure 13 The example of FIG shows a case where the base station 910 determines whether the RSRP satisfies a threshold value and decides whether to perform retraining accordingly. However, the UE 920 can determine whether the RSRP satisfies the threshold value. In other words, the entity that determines whether to perform retraining of the autoencoder can be the base station 910 or the UE 920.

[0359] When the base station 910 is the subject of determining whether to perform retraining of the autoencoder, the base station 910 may determine whether the RSRP value meets the threshold and decide whether to perform retraining. In this case, the base station 910 may determine retraining based on a comparison result between the RSRP value reported by the UE 920 and the RSRP threshold.

[0360] On the other hand, when the UE 920 is the subject of determining whether to perform retraining of the autoencoder, the UE 920 may determine retraining based on a comparison between the RSRP measured by the UE 920 and the RSRP threshold. In this case, when the RSRP threshold is determined by the base station 910, the RSRP threshold may be provided to the UE 920 in advance by the base station 910.

[0361] When the UE 920 is the subject that determines whether to perform retraining of the autoencoder, and the RSRP threshold is determined by the base station 910 , the base station 910 may use various signaling methods to provide the UE 920 with RSRP threshold information.

[0362] For example, the RSRP threshold information transmitted to the UE 920 may be transmitted using an SIB, or may be transmitted by being included in an RRC reconfiguration message or a MAC-CE message.

[0363] As another example, RSRP threshold information may be transmitted to the UE 920 via DCI.

[0364] As another example, the RSRP threshold information may be transmitted to the UE 920 via a second message (Msg2) of a four-step RACH procedure or a message B (MsgB) of a two-step RACH procedure.

[0365] As another example, if a new RRC signaling message is defined to send RSRP threshold information, the newly defined RRC signaling message may be used to send the RSRP threshold information to the UE 920 .

[0366] When the UE 920 is the subject of determining whether the autoencoder needs to be retrained, the UE 920 may notify the base station 910 that the autoencoder needs to be retrained. The UE 920 may notify the base station 910 of information indicating that the autoencoder needs to be retrained using various forms of signaling messages.

[0367] For example, the UE 920 may transmit information indicating retraining of the autoencoder to the base station 910 through UCI, the first message (Msg1) in the four-step RACH procedure, or message A (MsgA) in the two-step RACH procedure.

[0368] As another example, the UE 920 may use the UE assistance information to send information indicating that the autoencoder needs to be retrained to the base station 910 .

[0369] As another example, if a new RRC signaling message is defined for the UE 920 to send information indicating that the autoencoder needs to be retrained to the base station 910, the defined RRC signaling message may be used.

[0370] When the base station 910 is Figure 12 When it is determined in step S1200 or in a separate step that the subject responsible for determining whether the autoencoder needs to be retrained is UE 920, base station 910 can notify UE 920 accordingly. Notifying UE 920 that the subject responsible for determining the necessity of retraining the autoencoder is UE 920 can be understood as a request for UE 920 to determine whether the autoencoder needs to be retrained. Therefore, in this case, base station 910 can request retraining determination information from UE 920 using a UE Information Request message. UE 920 can then determine whether the autoencoder needs to be retrained, as described above, and can report information regarding whether the autoencoder needs to be retrained to base station 910 using a UE Information Response message.

[0371] When RSRP does not meet the expected value at step S1310, in other words, when the saturated RSRP is less than the RSRP threshold, the process may proceed to step S1320. When the saturated RSRP is greater than or equal to the RSRP threshold, the base station 910 may proceed to step S1320.

[0372] Since step S1320 corresponds to a situation where the saturated RSRP does not meet the desired RSRP (ie, the threshold RSRP value), the autoencoder needs to be retrained. The base station 910 may generate information related to the retraining.

[0373] In step S1330 , the base station 910 may send the generated information related to the retraining of the autoencoder to the UE 920 .

[0374] Examples of the information related to the retraining of the autoencoder generated in step S1320 and transmitted to the UE 920 in step S1330 may include the information shown in Table 9 below.

[0375] [Table 9]

[0376]

[0377] As shown in Table 9, the delivered information may include encoder retraining information indicating retraining of the encoder, encoder reconfiguration information, and an encoder identifier.

[0378] The encoder identifier (ID) exemplified in Table 9 may indicate information for notifying the UE 920 of which encoder is currently used for online training by the base station 910. The UE 920 may identify which encoder is indicated by the base station 910 through the encoder identifier received from the base station 910.

[0379] In addition, the encoder retraining information illustrated in Table 9 may indicate that retraining has started because the trained decoder does not meet a certain level of RSRP through the CFI transmitted by the UE 920. Therefore, the encoder retraining information may indicate re-execution of the training process of the encoder and decoder having the self-encoder identifier.

[0380] The encoder reconfiguration information in Table 9 may include information about the discarded nodes, as described above with reference to Figure 11 The information about the dropped nodes may include information about the nodes to be dropped among the nodes forming each layer of the encoder 9111 stored in the UE 920, and information about the drop probability of the corresponding nodes. Generally, among the nodes forming the encoder 9111, nodes in the hidden layer 1120 may be subject to dropout.

[0381] When the hidden layer 1120 is composed of one or more layers, all nodes in a specific hidden layer should not be discarded. This is because if all nodes in a specific hidden layer 1120 are discarded, the encoding operation cannot be performed. To prevent this, the present invention can define the minimum and maximum number of nodes that form a specific layer.

[0382] For example, it can be assumed that the hidden layer 1120 consists of two layers, the first hidden layer is connected to the input layer 1110, and the second hidden layer is connected to the output layer 1130. In addition, it can be assumed that the minimum number and the maximum number of nodes forming the first hidden layer and the second hidden layer are equal.

[0383] Then, the number of nodes formed in the first hidden layer by applying dropout can satisfy the following relationship, and the second hidden layer can also satisfy the same relationship.

[0384] <Relationship between the maximum number of nodes, the minimum number of nodes, and the number of nodes used for the hidden layer due to dropout probability>

[0385] The minimum number of nodes forming the first hidden layer ≤ the number of nodes in the first hidden layer formed by applying the dropout probability ≤ the maximum number of nodes forming the first hidden layer

[0386] The information about the dropped node may be determined based on the above relationship.

[0387] On the other hand, in the encoder identifier parameter, the suffix "n" can indicate the number of times that discarding has been applied. When discarding is not applied, the identifier can consist of only the encoder ID. In the case of an encoder that has applied discarding once, the identifier can be configured as "Encoder_ID_1", where "n" indicates the number of times discarding has been applied. Therefore, the value of "n" can increase according to the number of times discarding has been applied.

[0388] The scheme for transmitting the encoder reconfiguration information may not be configured in step S1320, but may be pre-configured. For example, when the base station 910 provides the encoder 9111 to the UE 920 after performing offline training, the base station 910 may additionally pre-transmit information about the order of nodes to be discarded for all nodes in each layer required for retraining the encoder. As another example, when the base station 910 provides the encoder 9111 to the UE 920 after performing offline training, the base station 910 may send information about the order of nodes to be discarded for each layer to the UE 920. The transmission of such information may be in Figure 12 Step S1200 is performed in advance, as described above.

[0389] As described above, when the base station 910 transmits information about the dropped nodes to the UE 920 in advance, the base station 910 may transmit only the encoder retraining information and the encoder identifier to the UE 920 when retraining is required. The UE 920 may then recognize from the encoder retraining parameters that encoder reconfiguration is required, and may apply the dropping to the corresponding node based on the previously agreed upon dropped node information.

[0390] When discard is indicated for the first time, the base station 910 may inform the UE 920 of the indication of initial discard by setting the encoder identifier to “Encoder_ID_1” and transmitting it to the UE 920 , as described above.

[0391] When the base station 910 transmits information related to autoencoder retraining, such as that shown in Table 9, to the UE 920, various forms of signaling may be used. For example, the information related to autoencoder retraining, such as that shown in Table 9, may be transmitted to the UE 920 via an SIB, or may be included in an RRC reconfiguration message or a MAC-CE message.

[0392] As another example, information related to the retraining of the autoencoder, such as Table 9, may be transmitted to the UE 920 via DCI.

[0393] As another example, the autoencoder retraining related information such as Table 9 may be transmitted to the UE 920 via the second message (Msg2) of the four-step RACH procedure or message B (MsgB) of the two-step RACH procedure.

[0394] As another example, if a new RRC signaling message is defined to transmit information related to autoencoder retraining, such as that in Table 9, the newly defined RRC signaling message may be used to transmit the information to the UE 920 .

[0395] In step S1330 , the UE 920 may receive information related to retraining of the autoencoder based on the above solution.

[0396] In step S1340, the UE 920 may reconfigure the autoencoder based on the information related to the retraining of the autoencoder. In the present invention, the component to be reconfigured by the UE 920 may be the encoder of the autoencoder. In other words, the operation may be Figure 11 The specific layer configuration of the encoder described in

[15] is dropped.

[0397] At step S1350, the UE 920 may transmit an encoder reconfiguration completion message to the base station 910 after the discarding in the encoder is completed. The encoder reconfiguration completion message may indicate that the discarding of the encoder for a specific autoencoder has been completed. Therefore, the encoder reconfiguration completion message may include the identifier information of the autoencoder. In other words, the UE 920 may report to the base station 910 which autoencoder has been reconfigured and how many times discarding has been applied. Through this, the base station 910 can confirm whether the reconfigured autoencoder is the autoencoder indicated by the base station 910 and whether the number of discarding applications is the same as indicated.

[0398] Since the present invention describes an example for downlink, the autoencoder reported by UE 920 may specifically be an encoder. In addition, when at least one of the nodes forming a specific layer of the encoder is dropped, the input to and output from the corresponding node may be set to a weight of "0." In other words, both the input and the output may be set to "0."

[0399] When the UE 920 sends the auto-encoder reconfiguration completion message to the base station 910 at step S1350 , the UE 920 may send the auto-encoder reconfiguration completion message to the base station 910 through various forms of signaling.

[0400] For example, the UE 920 may transmit the self-encoder reconfiguration completion information to the base station 910 by using UCI or UE assistance information.

[0401] As another example, the UE 920 may pre-report the self-encoder reconfiguration completion message to the base station 910 by utilizing the first message (Msg1) of the four-step RACH procedure or message A (MsgA) of the two-step RACH procedure.

[0402] As another example, if a new RRC signaling message is defined to transmit information requested from the base station 910 , the UE 920 may provide the requested information using the newly defined RRC signaling message.

[0403] As another example, the base station 910 may request transmission of the autoencoder reconfiguration complete message by sending a UE information request message to the UE 920. In this case, the UE 920 may report the autoencoder reconfiguration complete message by including the autoencoder reconfiguration complete message in a UE information response message to the base station 910.

[0404] In step S1350 , the base station 910 may receive a self-encoder reconfiguration completion message.

[0405] In step S1360, the base station 910 may perform an autoencoder retraining process based on the reception of the autoencoder reconfiguration completion message. Figure 12 The CSI-RS transmission steps described in [1] are restarted.

[0406] [C] Forced termination process of online training

[0407] On the other hand, when executing Figures 12 to 13 During the online training process, the time used to train the autoencoder through online training between the base station 910 and the UE 920 can be continuously increased. For example, if the expected RSRP is not met, the time used to train the autoencoder can continue to increase. When performing online training, traffic can be generated between the base station 910 and the UE 920, which can increase the congestion level of the wireless channel and the use of frequency resources. In addition, since the UE 920 is generally operated by a battery, the battery consumption of the UE 920 may also increase. Therefore, even if the expected RSRP is not met, it is preferable to terminate the training or retraining process at an appropriate time.

[0408] The present invention describes a process for configuring online training expiration information and terminating online training when the online training expires.

[0409] Figure 15 is a sequence diagram for describing a process for terminating online training of an autoencoder at a base station.

[0410] In the description Figure 15 Previously, it was assumed that the base station and UE correspond to Figures 9 to 13 Therefore, it should be noted that the base station 910 and the UE 920 described in Figure 9 The same reference numerals as those described in are used for the UE 920 and the base station 910.

[0411] In step S1500, the base station 910 may initiate online training of the autoencoder. The initiation of online training of the autoencoder may be based on Figures 12 to 13 The initial online training process is performed using the procedure described in

[15] .

[0412] In step S1502, the base station 910 may obtain an RSRP value during the online training process of the autoencoder and may store the obtained RSRP value and information about the autoencoder corresponding thereto. Here, the RSRP value may be obtained based on the CFI reported by the UE 920, as described above.

[0413] In step S1504, the base station 910 may determine whether the acquired RSRP satisfies the expected RSRP value. In other words, step S1504 may be a step for determining whether the RSRP exceeds the expected RSRP value. Figure 12 When the obtained RSRP value meets the expected RSRP value, the base station 910 may terminate the online training of the autoencoder.

[0414] On the other hand, when the RSRP value obtained in step S1504 does not meet the expected RSRP value, the base station 910 may proceed to step S1506 to determine whether the condition for forced termination of online training is met. Here, the condition for forced termination of online training may be defined by various values.

[0415] An example of such a forced termination condition is as follows.

[0416] For example, by defining the maximum time allocated to online training as "OnlineTrainingTimeMax", it is possible to determine whether the mandatory termination condition is satisfied by checking whether the set time has passed.

[0417] As another example, a maximum number of retraining iterations for forced termination of online training may be set, and the base station 910 may determine whether the forced termination condition is satisfied by checking whether the maximum number of retraining iterations has been reached. Hereinafter, for ease of description, a method for determining whether the forced termination condition is satisfied based on the maximum time allocated to online training will be described.

[0418] The value of OnlineTrainingTimeMax may be set by the base station 910 or may be set by the UE 920 based on its battery capacity or hardware requirements. If the UE 920 sets the OnlineTrainingTimeMax value, the UE 920 may send the OnlineTrainingTimeMax value to the base station 910 before or at the initiation of online training of the autoencoder so that the base station 910 can check for time expiration.

[0419] As another example of UE 920 determining the OnlineTrainingTimeMax value, base station 910 may send a group of possible OnlineTrainingTimeMax values ​​to UE 920 according to the situation of UE 920, and UE 920 may select a value from the group and send the selected value to base station 910.

[0420] The OnlineTrainingTimeMax value may be set based on various factors such as the channel environment between the base station 910 and the UE 920 or the mobility of the UE 920. For example, in an environment where the channel changes rapidly or the mobility of the UE 920 is high, the base station 910 may set a smaller OnlineTrainingTimeMax value to achieve faster training for more accurate channel estimation and faster adaptation to channel changes.

[0421] When the UE 920 transmits the OnlineTrainingTimeMax value set by the UE 920 to the base station 910, the UE 920 may use various forms of signaling. For example, the UE 920 may transmit the set OnlineTrainingTimeMax value to the base station 910 via UCI or UE assistance information. As another example, the UE 920 may transmit the set OnlineTrainingTimeMax value to the base station 910 via the first message (Msg1) of the four-step RACH procedure or message A (MsgA) of the two-step RACH procedure.

[0422] If the base station 910 requests the UE 920 to report information about the maximum time allocable for online training as an OnlineTrainingTimeMax value through the UE Information Request message, the UE 920 may respond with a UE Information Response message to report the maximum time allocable for online training.

[0423] When at least one of the above solutions is used to configure a mandatory termination condition for online training, the base station 910 may determine whether the mandatory termination condition is satisfied in step S1506. In other words, the base station 910 may determine whether the mandatory termination condition is satisfied by checking whether a timer set to OnlineTrainingTimeMax has expired.

[0424] If the forced termination condition is satisfied at step S1506, the base station 910 may proceed to step S1508. If the forced termination condition is not satisfied, the base station 910 may proceed to step S1512.

[0425] When proceeding to step S1508 , the base station 910 may determine a model having the highest RSRP value among the autoencoder models trained through online training up to that point as a model for downlink channel estimation.

[0426] At step S1510, the base station 910 may transmit the determined channel estimation model to the UE 920. In this case, the information about the channel estimation model may be an encoder identifier corresponding thereto. In other words, the encoder identifier information transmitted to the UE 920 may indicate a specific autoencoder among a plurality of autoencoders and, as described above, may include a drop count indicating how many times drop has been applied.

[0427] An example of the information transmitted from the base station 910 to the UE 920 at step S1510 is shown in Table 10 below.

[0428] [Table 10]

[0429]

[0430] The training expiration information exemplified in Table 10 may indicate (or notify) UE 920 that online training of the autoencoder has terminated. Furthermore, the encoder identifier may indicate a specific autoencoder (or an encoder of the autoencoders) among the multiple autoencoders, as described above. The suffix "3" may indicate that three discards have been performed. Thus, the best encoder can be selected from the encoders that have undergone online training up to that point.

[0431] In addition, the training expiration information and encoder information (e.g., those shown in Table 10) can be transmitted from the base station 910 to the UE 920 via various forms of signaling, as described above. For example, the training expiration information and encoder information (e.g., Table 10) can be transmitted from the base station 910 to the UE 920 via one of an SIB, a MAC-CE, or an RRC reconfiguration message. As another example, the training expiration information and encoder information (e.g., Table 10) can be transmitted from the base station 910 to the UE 920 via a DCI. As another example, the training expiration information and encoder information (e.g., Table 10) can be transmitted from the base station 910 to the UE 920 via the second message (Msg2) of the four-step RACH procedure or message B (MsgB) of the two-step RACH procedure. As another example, when a new RRC signaling message is defined to transmit the training expiration information and encoder information (e.g., Table 10), the training expiration information and encoder information can be transmitted from the base station 910 to the UE 920 using the newly defined RRC signaling message.

[0432] On the other hand, when it is determined in step S1506 that the conditions for forced termination of online training are not satisfied, the base station 910 may proceed to step S1512 to perform an online retraining process. The online retraining process may include Figure 12 Steps S1250 to S1270 described in Figure 13 Steps S1300 to S1350 described in .

[0433] For each iteration, the base station 910 may store the RSRP value obtained through online retraining, the corresponding model information, and the discard count information. By doing so, the base station 910 may identify the RSRP value having the highest value among the RSRP values ​​obtained through online training of the autoencoder in step S1508, and may obtain the identifier of the autoencoder having the highest RSRP value and its discard count information in step S1510.

[0434] [D] Normal termination process of online training

[0435] In the following, the process of normal termination of online training will be described. This can correspond to Figure 13 In the case of step S1310, the base station 910 determines that the saturated RSRP value meets the expected RSRP value.

[0436] although Figure 13 This is merely described as a form of online training termination, and the base station 910 may notify the UE 920 of the termination of the online training. Table 11 below shows an example of the base station 910 sending information to the UE 920 to notify the termination of the online training.

[0437] [Table 11]

[0438] Delivered information Base station → UE Encoder training completed

[0439] As shown in Table 11, base station 910 may send encoder training completion indication information to UE 920. The encoder training completion information may indicate that the encoder trained at the current time is ready for use. Therefore, after receiving the encoder training completion information, UE 920 may update the encoder to the encoder that was in the online training state before receiving the encoder training completion information. Furthermore, base station 910 may also store the encoder and decoder of the self-encoder in corresponding states based on the encoder training completion information.

[0440] The encoder training completion information may be transmitted from the base station 910 to the UE 920 via various forms of signaling. For example, the encoder training completion information (e.g., Table 11) may be transmitted from the base station 910 to the UE 920 via one of an SIB, a MAC-CE, or an RRC reconfiguration message. As another example, the encoder training completion information (e.g., Table 11) may be transmitted from the base station 910 to the UE 920 via a DCI. As another example, the encoder training completion information (e.g., Table 11) may be transmitted from the base station 910 to the UE 920 via the second message (Msg2) of a four-step RACH procedure or message B (MsgB) of a two-step RACH procedure. As another example, if a new RRC signaling message is defined for transmitting the encoder training completion information (e.g., Table 11), the encoder training completion information may be transmitted from the base station 910 to the UE 920 using the newly defined RRC signaling message.

[0441] On the other hand, thereafter, the UE 920 and the base station 910 may perform model inference using the obtained autoencoder model for downlink channel estimation transmitted from the base station 910 based on the parameters obtained in the current online training process.

[0442] The model inference process may be a step of performing actual channel estimation and may follow a process similar to the online training process. In other words, the base station 910 may transmit a CSI-RS, and the UE 920 may obtain a CFI from the received CSI-RS through an encoder and report the obtained CFI to the base station 910. The base station 910 may then obtain a received CSI-RS based on the received CFI through a decoder and obtain an RSRP value, thereby determining the downlink CSI.

[0443] The content of each of the above sections can be applied through simple combination, partial combination or extended combination.

[0444] Figure 16 It is a conceptual diagram for describing a case where the overall operation of the present invention is combined and executed.

[0445] Step S1610 may be an operation performed at the base station 910, and step S1611 may correspond to Figure 12 In step S1210 described in . In step S1611 , the base station 910 may determine a CFI transmission period based on the network size and the channel coherence time.

[0446] Step S1610 may correspond to Figure 12 In step S1612, the base station 910 may determine the LVD based on the channel resolution and delay requirements.

[0447] Step S1620 may correspond to Figure 12 In step S1620 , the base station 910 may transmit the CFI transmission period and LVD to the UE 920 , and transmit the CSI-RS to the UE 920 .

[0448] Step S1630 may correspond to Figure 12 In step S1630 , the UE 920 may determine the CFI dimension, compress the received CSI-RS through the encoder, and report the CFI to the base station 910 .

[0449] Step S1640 may correspond to Figure 13 In step S1640, when the RSRP is saturated, the base station 910 may stop training and compare the saturated RSRP value with the RSRP threshold.

[0450] When the saturated RSRP value of step S1640 is greater than or equal to the RSRP threshold, step S1650 may be performed, and when the saturated RSRP value of step S1640 is less than the RSRP threshold, step S1660 may be performed.

[0451] Step S1650 may perform the operations described in Section [D]. In other words, at step S1650, the base station 910 may send training completion information to the UE 920 after the training is completed.

[0452] Step S1660 may correspond to Figure 13 Therefore, at step S1660, the base station 910 may send the retraining information, the encoder reconfiguration information, and the encoder identifier information to the UE 920.

[0453] Step S1670 may correspond to Figure 13 Therefore, at step S1670, the UE 920 may reconfigure the encoder based on the retraining information, the encoder reconfiguration information, and the encoder identifier information, and report encoder reconfiguration completion information to the base station 910.

[0454] In step S1670, if the training is completed within the allocated time, the UE 920 may report encoder reconfiguration completion information to the base station 910 after reconfiguring the encoder. If the training is not completed within the allocated time in step S1670, step S1680 may be performed.

[0455] Step S1680 may correspond to a situation where the training is not completed within the allotted time, such as in Sections [C] and Figure 15 Therefore, in step S1680, the base station 910 may send the training expiration information and the identifier of the best encoder to the UE 920.

[0456] above Figure 16 The case where all the processes described in the present invention are configured as one process is shown. However, it can be omitted or modified Figure 16 Some of these modifications have been described in the above figures. Since it is impossible to cover all variations in the present invention, it should be noted that various types of modifications and combinations can be made based on the content described in the present invention.

[0457] The operation of the method according to the exemplary embodiment of the present invention can be implemented as a computer-readable program or code in a computer-readable recording medium. The computer-readable recording medium may include all types of recording devices storing data that can be read by a computer system. In addition, the computer-readable recording medium can store and execute programs or codes, which can be distributed among computer systems connected via a network and read by computers in a distributed manner.

[0458] Computer readable recording media may include hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, or flash memory. Program instructions may include not only machine language codes created by a compiler, but also high-level language codes that can be executed by a computer using an interpreter.

[0459] Although some aspects of the present invention have been described in the context of an apparatus, these aspects can be indicated according to the corresponding description of the method, and a block or apparatus can correspond to the steps of the method or the features of the steps. Similarly, the aspects described in the context of the method can be represented as the features of the corresponding blocks or items or corresponding apparatus. Some or all steps of the method can be performed by (or using) a hardware device such as a microprocessor, a programmable computer or an electronic circuit. In some embodiments, one or more of the most important steps of the method can be performed by such an apparatus.

[0460] In some exemplary embodiments, a programmable logic device, such as a field programmable gate array, can be used to perform some or all of the functions of the methods described herein. In some exemplary embodiments, the field programmable gate array can be operated with a microprocessor to perform one of the methods described herein. Typically, the methods are preferably performed by specific hardware devices.

[0461] The description of the present invention is merely exemplary in nature, and therefore variations that do not depart from the essence of the present invention are intended to be within the scope of the present invention. Such variations should not be considered to depart from the spirit and scope of the present invention. Therefore, it will be understood by those skilled in the art that various changes in form and details may be made without departing from the spirit and scope defined by the appended claims.

Claims

1. A method of a user equipment (UE), comprising: receiving, from a base station, a channel characteristic indicator (CFI) transmission period and a latent variable dimension (LVD) of an encoder performing online learning; determining a first encoder to perform online learning based on the CFI transmission period and the LVD; receiving a first reference signal (RS) from a base station; generating a first CFI by compressing the received first RS by a first encoder; as well as Sending a first CFI to the base station based on a CFI transmission period, The first CFI is determined based on the product of the number of nodes in the latent space of the first encoder and the LVD.

2. The method according to claim 1, wherein The LVD corresponds to the number of bits required to represent each latent variable value included in the latent space of the first encoder.

3. The method according to claim 1, wherein The CFI transmission period is determined based on at least one of a network size or a channel coherence time of the first encoder.

4. The method according to claim 1, wherein The LVD is determined based on at least one of a channel resolution or a latency requirement of the base station.

5. The method according to claim 1, further comprising: receiving information related to retraining of the first encoder from the base station; reconfiguring the first encoder based on the retraining-related information; receiving a second RS from a base station; generating a second CFI by compressing the received second RS by the reconfigured first encoder; as well as The second CFI is sent to the base station.

6. The method according to claim 5, wherein: The retraining-related information includes at least one of information indicating retraining of the encoder, information on a dropped node of the encoder, or identifier information of the first encoder, and the identifier information of the first encoder further includes a drop count.

7. The method according to claim 1, further comprising: receiving, from the base station, training termination information indicating termination of training of the first encoder; before receiving the training termination information, updating the first encoder based on information obtained by training the first encoder; receiving a third RS from the base station; generating a third CFI by compressing the third RS by the updated first encoder; as well as The third CFI is sent to the base station.

8. The method according to claim 1, further comprising: receiving training expiration information including optimal encoder information from a base station; updating the first encoder based on the best encoder information from the base station; configuring the updated first encoder as a channel estimation encoder; receiving a fourth RS from the base station; generating a fourth CFI by compressing a fourth RS by the updated first encoder; as well as Sending a fourth CFI to the base station, The optimal encoder information includes an encoder identifier and information about the number of discards of the encoder.

9. A base station method, comprising: determining a channel characteristic indicator (CFI) transmission period and a latent variable dimension (LVD) of an encoder performing online learning with a user equipment (UE); Send CFI transmission period and LVD to UE; receiving a first CFI from the UE based on a CFI transmission period; as well as A received reference signal (RS) is restored from the received first CFI by the decoder.

10. The method according to claim 9, further comprising: Obtaining a first reference signal received power (RSRP) value of the recovered received RS; Check whether the first RSRP value is saturated; Based on the saturation of the first RSRP value, checking whether the first RSRP value is equal to or greater than a preset threshold; as well as Based on the first RSRP value being equal to or greater than a preset threshold, sending encoder training termination information to the UE, Here, the first CFI is determined based on the product of the number of nodes in the latent space of the encoder performing online learning and the LVD.

11. The method according to claim 10, wherein: The first RSRP value is determined to be saturated based on the first RSRP value being equal to or within a predetermined range of an RSRP value obtained from a CFI previously received by the UE.

12. The method according to claim 11, further comprising: generating information related to retraining of the encoder based on that the first RSRP value is not saturated; Send information related to encoder retraining to UE; as well as The encoder retraining process is performed by the UE.

13. The method according to claim 12, wherein: The retraining-related information includes at least one of information indicating retraining of the encoder, information on a dropped node of the encoder, or identifier information of the encoder, and the identifier information of the encoder further includes a drop count.

14. The method according to claim 13, further comprising: checking a number of drops corresponding to an RSRP having a highest RSRP value among the trained RSRP values ​​based on the retraining not being completed within a predetermined time; as well as The identifier information of the encoder including the number of discarding times and the training expiration information are sent to the UE.

15. The method according to claim 9, wherein The CFI transmission period is determined based on at least one of a network size or a channel coherence time of an encoder performing online training.

16. The method according to claim 9, wherein The LVD is determined based on at least one of a channel resolution or a latency requirement of the base station.

17. The method according to claim 9, further comprising: Sending a second RS to the UE after training is completed; receiving, from the UE, a second CFI corresponding to a second RS; Obtaining a second RSRP value of the received RS from the second CFI; as well as A downlink channel toward the UE is estimated using the second RSRP value.

18. A user equipment (UE), comprising: a transceiver configured to transmit signals to and receive signals from a base station; as well as at least one processor, The at least one processor causes the UE to execute: receiving, from a base station, a channel characteristic indicator (CFI) transmission period and a latent variable dimension (LVD) of an encoder performing online learning; determining a first encoder to perform online learning based on the CFI transmission period and the LVD; receiving a first reference signal (RS) from a base station; generating a first CFI by compressing the received first RS by a first encoder; and Sending a first CFI to the base station based on a CFI transmission period, The first CFI is determined based on the product of the number of nodes in the latent space of the first encoder and the LVD.

19. The UE according to claim 18, wherein: The CFI transmission period is determined based on at least one of a network size or a channel coherence time of the first encoder, and the LVD is determined based on at least one of a channel resolution or a delay requirement of the base station.

20. The UE according to claim 18, wherein The at least one processor further causes the UE to perform: receiving information related to retraining of the first encoder from the base station; reconfiguring the first encoder based on the retraining-related information; receiving a second RS from a base station; generating a second CFI by compressing the received second RS by the reconfigured first encoder; as well as Sending a second CFI to the base station, The retraining-related information includes at least one of information indicating retraining of the encoder, information about a discarded node of the encoder, or identifier information of the encoder, and the identifier information of the first encoder further includes a discarding count.

Images